DavidAU/Qwen3.5-9B-The-Deckard-Pro-Writer-4Banger-Uncensored-Heretic

Qwen3.5-9B-The-Deckard-Pro-Writer-4Banger-Uncensored-Heretic

The Anti-Slop retraining project with Deckard STRICT training and "ANDR" dataset.

This model is a test case to apply (these methods) to other Qwen(s) and other model arches (Gemma, Mistral, Llama, etc etc).

Be gone "not this and that" prose, "suddenly" and "semi-colon" hell.

Kill the SLOP.

3 Example outputs below, from Test Quant, Q4_K_S, non-imatrix.

Testing shows it kills 99% of the "Qwen Prose" (so far, may change) issues and also addresses these issues in the reasoning/thinking block too.

ANDR dataset is very specific... 4 banger uses 4 novels as a base to FOCUS THE MODEL.

This is also a HERETIC model, trained post "Heretic'ing" -> this model does what you want, no questions asked.

Fully uncensored.

VISION:

• Vision (images) tested -> working with new training.

---

DE-CENSORING:

Performance

KLD of less than 1 is excellent, zero is perfect.

Metric	This model	Original model (Qwen/Qwen3.5-9B)
KL divergence	0.0793	0 (by definition)
Refusals	6/100	100/100

NOTES:

• Suggest min q4ks (non-imatrix) or IQ3S (imatrix).
• Tested with rep pen of 1 (off).
• Context: 256k (default).

IMPORTANT:

• Other versions in testing.
• Information from Qwen's repo below.
• Video portions of the model were NOT TESTED.

---

Using an "uncensored" (refusals removed) model VS trained "uncensored" model

Usually when you a tell a model to generate horror, swear or x-rated content this is all you have to do to get said content type.

In the case of this model, it will not refuse your request, however it needs to be "pushed" a bit / directed a bit more in SOME CASES.

Although this model will generated x-rated content too, likewise you need to tell it to use "slang" (and include the terms you want) to get it generate the content correctly as the "expected" content level too.

Without these added directive(s), the content can be "bland" by comparison to an "uncensored model" or model trained on uncensored content.

Roughly, the model tries to generate the content but the "default" setting(s) are so "tame" it needs a push to generate at expected graphic, cursing or explicit levels.

Even with minimal direction (ie, use these words to swear: x,y,z), this will be enough to push the model to generate the requested content in the ahh... expected format.

---

Settings: CHAT / ROLEPLAY and/or SMOOTHER operation of this model:

In "KoboldCpp" or "oobabooga/text-generation-webui" or "Silly Tavern" ;

Set the "Smoothing_factor" to 1.5

: in KoboldCpp -> Settings->Samplers->Advanced-> "Smooth_F"

: in text-generation-webui -> parameters -> lower right.

: In Silly Tavern this is called: "Smoothing"

NOTE: For "text-generation-webui"

-> if using GGUFs you need to use "llama_HF" (which involves downloading some config files from the SOURCE version of this model)

Source versions (and config files) of my models are here:

https://huggingface.co/collections/DavidAU/d-au-source-files-for-gguf-exl2-awq-gptq-hqq-etc-etc-66b55cb8ba25f914cbf210be

OTHER OPTIONS:

• Increase rep pen to 1.1 to 1.15 (you don't need to do this if you use "smoothing_factor")

• If the interface/program you are using to run AI MODELS supports "Quadratic Sampling" ("smoothing") just make the adjustment as noted.

Highest Quality Settings / Optimal Operation Guide / Parameters and Samplers

This a "Class 1" model:

For all settings used for this model (including specifics for its "class"), including example generation(s) and for advanced settings guide (which many times addresses any model issue(s)), including methods to improve model performance for all use case(s) as well as chat, roleplay and other use case(s) please see:

[ https://huggingface.co/DavidAU/Maximizing-Model-Performance-All-Quants-Types-And-Full-Precision-by-Samplers_Parameters ]

You can see all parameters used for generation, in addition to advanced parameters and samplers to get the most out of this model here:

[ https://huggingface.co/DavidAU/Maximizing-Model-Performance-All-Quants-Types-And-Full-Precision-by-Samplers_Parameters ]

---

Qwen3.5-9B

This repository contains model weights and configuration files for the post-trained model in the Hugging Face Transformers format.

These artifacts are compatible with Hugging Face Transformers, vLLM, SGLang, KTransformers, etc.

Over recent months, we have intensified our focus on developing foundation models that deliver exceptional utility and performance. Qwen3.5 represents a significant leap forward, integrating breakthroughs in multimodal learning, architectural efficiency, reinforcement learning scale, and global accessibility to empower developers and enterprises with unprecedented capability and efficiency.

Qwen3.5 Highlights

Qwen3.5 features the following enhancement:

• Unified Vision-Language Foundation: Early fusion training on multimodal tokens achieves cross-generational parity with Qwen3 and outperforms Qwen3-VL models across reasoning, coding, agents, and visual understanding benchmarks.

• Efficient Hybrid Architecture: Gated Delta Networks combined with sparse Mixture-of-Experts deliver high-throughput inference with minimal latency and cost overhead.

• Scalable RL Generalization: Reinforcement learning scaled across million-agent environments with progressively complex task distributions for robust real-world adaptability.

• Global Linguistic Coverage: Expanded support to 201 languages and dialects, enabling inclusive, worldwide deployment with nuanced cultural and regional understanding.

• Next-Generation Training Infrastructure: Near-100% multimodal training efficiency compared to text-only training and asynchronous RL frameworks supporting massive-scale agent scaffolds and environment orchestration.

For more details, please refer to our blog post Qwen3.5.

Model Overview

• Type: Causal Language Model with Vision Encoder
• Training Stage: Pre-training & Post-training
• Language Model
  • Number of Parameters: 9B
  • Hidden Dimension: 4096
  • Token Embedding: 248320 (Padded)
  • Number of Layers: 32
  • Hidden Layout: 8 × (3 × (Gated DeltaNet → FFN) → 1 × (Gated Attention → FFN))
  • Gated DeltaNet:
    • Number of Linear Attention Heads: 32 for V and 16 for QK
    • Head Dimension: 128
  • Gated Attention:
    • Number of Attention Heads: 16 for Q and 4 for KV
    • Head Dimension: 256
    • Rotary Position Embedding Dimension: 64
  • Feed Forward Network:
    • Intermediate Dimension: 12288
  • LM Output: 248320 (Padded)
  • MTP: trained with multi-steps
• Context Length: 262,144 natively and extensible up to 1,010,000 tokens.

Benchmark Results

Language

	GPT-OSS-120B	GPT-OSS-20B	Qwen3-Next-80B-A3B-Thinking	Qwen3-30BA3B-Thinking-2507	Qwen3.5-9B	Qwen3.5-4B
Knowledge & STEM
MMLU-Pro	80.8	74.8	82.7	80.9	82.5	79.1
MMLU-Redux	91.0	87.8	92.5	91.4	91.1	88.8
C-Eval	76.2	71.4	89.7	87.4	88.2	85.1
SuperGPQA	54.6	48.5	60.8	56.8	58.2	52.9
GPQA Diamond	80.1	71.5	77.2	73.4	81.7	76.2
Instruction Following
IFEval	88.9	88.2	88.9	88.9	91.5	89.8
IFBench	69.0	65.1	61.5	51.5	64.5	59.2
MultiChallenge	45.3	40.1	51.3	46.5	54.5	49.0
Long Context
AA-LCR	50.7	30.7	51.7	49.0	63.0	57.0
LongBench v2	48.2	45.6	48.0	44.8	55.2	50.0
Reasoning & Coding
HMMT Feb 25	90.0	76.7	73.7	63.1	83.2	74.0
HMMT Nov 25	90.0	81.8	81.2	73.8	82.9	76.8
LiveCodeBench v6	82.7	74.6	68.7	66.0	65.6	55.8
OJBench	41.5	36.3	29.7	25.1	29.2	24.1
General Agent
BFCL-V4	--	--	49.7	42.4	66.1	50.3
TAU2-Bench	--	--	57.4	41.9	79.1	79.9
VITA-Bench	--	--	29.5	14.1	29.8	22.0
DeepPlanning	--	--	0.4	4.9	18.0	17.6
Multilingualism
MMMLU	78.2	69.7	81.3	78.4	81.2	76.1
MMLU-ProX	74.5	67.3	73.6	69.1	76.3	71.5
NOVA-63	51.1	48.7	53.3	52.5	55.9	54.3
INCLUDE	74.0	65.3	78.3	74.4	75.6	71.0
Global PIQA	84.1	79.8	83.5	80.2	83.2	78.9
PolyMATH	54.0	30.9	62.4	52.6	57.3	51.1
WMT24++	74.4	67.8	57.4	69.3	72.6	66.6
MAXIFE	83.7	80.1	79.9	77.4	83.4	78.0

* TAU2-Bench: we follow the official setup except for the airline domain, where all models are evaluated by applying the fixes proposed in the Claude Opus 4.5 system card.

* MMLU-ProX: we report the averaged accuracy on 29 languages.
* WMT24++: a harder subset of WMT24 after difficulty labeling and rebalancing; we report the averaged scores on 55 languages using XCOMET-XXL.
* MAXIFE: we report the accuracy on English + multilingual original prompts (totally 23 settings).
* Empty cells (--) indicate scores not yet available or not applicable.

Vision Language

	GPT-5-Nano-2025-08-07	Gemini-2.5-Flash-Lite	Qwen3-VL-30B-A3B	Qwen3.5-9B	Qwen3.5-4B
STEM and Puzzle
MMMU	75.8	73.4	76.0	78.4	77.6
MMMU-Pro	57.2	59.7	63.0	70.1	66.3
MathVision	62.2	52.1	65.7	78.9	74.6
Mathvista(mini)	71.5	72.8	81.9	85.7	85.1
We-Math	62.5	32.1	70.0	75.2	75.4
DynaMath	78.0	69.9	80.1	83.6	83.3
ZEROBench	1.0	1.0	0.0	3.0	3.0
ZEROBench_sub	22.2	19.2	23.7	31.1	26.3
VlmsAreBlind	66.7	68.4	72.5	93.7	92.6
BabyVision	14.4	17.5	18.6	28.6/25.8	16.0/19.1
General VQA
RealWorldQA	71.8	72.2	77.4	80.3	79.5
MMStar	68.6	69.1	75.5	79.7	78.3
MMBenchEN-DEV-v1.1	80.3	82.7	88.9	90.1	89.4
SimpleVQA	46.0	54.1	54.3	51.2	43.4
HallusionBench	58.4	64.5	66.0	69.3	65.0
Text Recognition and Document Understanding
OmniDocBench1.5	55.9	79.4	86.8	87.7	86.2
CharXiv(RQ)	50.1	56.1	56.6	73.0	70.8
MMLongBench-Doc	31.8	46.5	47.4	57.7	54.2
CC-OCR	58.9	72.9	77.8	79.3	76.7
AI2D_TEST	81.9	85.7	86.9	90.2	89.6
OCRBench	75.3	82.5	83.9	89.2	85.0
Spatial Intelligence
ERQA	45.8	44.3	45.3	55.5	54.0
CountBench	80.0	79.2	90.0	97.2	96.3
RefCOCO(avg)	--	--	89.3	89.7	88.1
EmbSpatialBench	74.2	66.1	80.6	83.0	81.3
RefSpatialBench	12.6	11.2	54.2	58.5	54.6
LingoQA	57.0	17.8	62.0	80.4	74.4
Hypersim	--	--	11.4	13.5	12.5
Nuscene	--	--	10.3	11.8	9.9
Video Understanding
VideoMME(w sub.)	71.7	74.6	79.9	84.5	83.5
VideoMME(w/o sub.)	66.2	72.7	73.3	78.4	76.9
VideoMMMU	63.0	69.2	75.0	78.9	74.1
MLVU	69.2	78.5	78.9	84.4	82.8
MVBench	--	--	72.0	74.4	71.2
LVBench	--	60.9	59.2	70.0	66.4
MMVU	63.1	65.3	66.1	67.8	64.9
Visual Agent
ScreenSpot Pro	--	--	60.5	65.2	60.3
OSWorld-Verified	--	--	30.6	41.8	35.6
AndroidWorld	--	--	55.0	57.8	58.6
Tool Calling
TIR-Bench	18.5	21.5	22.5	45.6/31.9	38.9/29.9
V*	68.1	69.6	83.2	90.1/88.5	84.3/86.4
Medical VQA
SLAKE	57.0	65.0	68.8	79.0	76.1
PMC-VQA	37.8	48.8	51.5	57.9	55.5
MedXpertQA-MM	26.7	35.3	35.5	49.9	42.9

* MathVision: our model’s score is evaluated using a fixed prompt, e.g., “Please reason step by step, and put your final answer within \boxed{}.” For other models, we report the higher score between runs with and without the \boxed{} formatting.
* BabyVision: scores reported as "with CI / without CI".
* TIR-Bench and V*: scores reported as "with CI / without CI".
* Empty cells (--) indicate scores not yet available or not applicable.

Quickstart

Qwen3.5 models operate in thinking mode by default, generating thinking content signified by <think>\n...</think>\n\n before producing the final responses. To disable thinking content and obtain direct response, refer to the examples here.

For streamlined integration, we recommend using Qwen3.5 via APIs. Below is a guide to use Qwen3.5 via OpenAI-compatible API.

Serving Qwen3.5

Qwen3.5 can be served via APIs with popular inference frameworks. In the following, we show example commands to launch OpenAI-Compatible API servers for Qwen3.5 models.

Inference efficiency and throughput vary significantly across frameworks. We recommend using the latest framework versions to ensure optimal performance and compatibility. For production workloads or high-throughput scenarios, dedicated serving engines such as SGLang, KTransformers or vLLM are strongly recommended.

The model has a default context length of 262,144 tokens. If you encounter out-of-memory (OOM) errors, consider reducing the context window. However, because Qwen3.5 leverages extended context for complex tasks, we advise maintaining a context length of at least 128K tokens to preserve thinking capabilities.

SGLang

SGLang is a fast serving framework for large language models and vision language models. SGLang from the main branch of the open-source repository is required for Qwen3.5, which can be installed using the following command in a fresh environment:

uv pip install 'git+https://github.com/sgl-project/sglang.git#subdirectory=python&egg=sglang[all]'

See its documentation for more details.

The following will create API endpoints at http://localhost:8000/v1:

• Standard Version: The following command can be used to create an API endpoint with maximum context length 262,144 tokens using tensor parallel on 8 GPUs.

  python -m sglang.launch_server --model-path Qwen/Qwen3.5-9B --port 8000 --tp-size 1 --mem-fraction-static 0.8 --context-length 262144 --reasoning-parser qwen3
  

• Tool Use: To support tool use, you can use the following command.

  python -m sglang.launch_server --model-path Qwen/Qwen3.5-9B --port 8000 --tp-size 1 --mem-fraction-static 0.8 --context-length 262144 --reasoning-parser qwen3 --tool-call-parser qwen3_coder
  

• Multi-Token Prediction (MTP): The following command is recommended for MTP:

  python -m sglang.launch_server --model-path Qwen/Qwen3.5-9B --port 8000 --tp-size 1 --mem-fraction-static 0.8 --context-length 262144 --reasoning-parser qwen3 --speculative-algo NEXTN --speculative-num-steps 3 --speculative-eagle-topk 1 --speculative-num-draft-tokens 4
  

vLLM

vLLM is a high-throughput and memory-efficient inference and serving engine for LLMs. vLLM from the main branch of the open-source repository is required for Qwen3.5, which can be installed using the following command in a fresh environment:

uv pip install vllm --torch-backend=auto --extra-index-url https://wheels.vllm.ai/nightly

See its documentation for more details.

For detailed Qwen3.5 usage guide, see the vLLM Qwen3.5 recipe.

The following will create API endpoints at http://localhost:8000/v1:

• Standard Version: The following command can be used to create an API endpoint with maximum context length 262,144 tokens using tensor parallel on 8 GPUs.

  vllm serve Qwen/Qwen3.5-9B --port 8000 --tensor-parallel-size 1 --max-model-len 262144 --reasoning-parser qwen3 
  

• Tool Call: To support tool use, you can use the following command.

  vllm serve Qwen/Qwen3.5-9B --port 8000 --tensor-parallel-size 1 --max-model-len 262144 --reasoning-parser qwen3 --enable-auto-tool-choice --tool-call-parser qwen3_coder 
  

• Multi-Token Prediction (MTP): The following command is recommended for MTP:

  vllm serve Qwen/Qwen3.5-9B --port 8000 --tensor-parallel-size 1 --max-model-len 262144 --reasoning-parser qwen3 --speculative-config '{"method":"qwen3_next_mtp","num_speculative_tokens":2}'
  

• Text-Only: The following command skips the vision encoder and multimodal profiling to free up memory for additional KV cache:

  vllm serve Qwen/Qwen3.5-9B --port 8000 --tensor-parallel-size 1 --max-model-len 262144 --reasoning-parser qwen3 --language-model-only
  

KTransformers

KTransformers is a flexible framework for experiencing cutting-edge LLM inference optimizations with CPU-GPU heterogeneous computing. For running Qwen3.5 with KTransformers, see the KTransformers Deployment Guide.

Hugging Face Transformers

Hugging Face Transformers contains a lightweight server which can be used for quick testing and moderate load deployment. The latest transformers is required for Qwen3.5:

pip install "transformers[serving] @ git+https://github.com/huggingface/transformers.git@main"

See its documentation for more details. Please also make sure torchvision and pillow are installed.

Then, run transformers serve to launch a server with API endpoints at http://localhost:8000/v1; it will place the model on accelerators if available:

transformers serve --force-model Qwen/Qwen3.5-9B --port 8000 --continuous-batching

Using Qwen3.5 via the Chat Completions API

The chat completions API is accessible via standard HTTP requests or OpenAI SDKs. Here, we show examples using the OpenAI Python SDK.

Before starting, make sure it is installed and the API key and the API base URL is configured, e.g.:

pip install -U openai

# Set the following accordingly
export OPENAI_BASE_URL="http://localhost:8000/v1"
export OPENAI_API_KEY="EMPTY"

We recommend using the following set of sampling parameters for generation

• Thinking mode for general tasks: temperature=1.0, top_p=0.95, top_k=20, min_p=0.0, presence_penalty=1.5, repetition_penalty=1.0
• Thinking mode for precise coding tasks (e.g. WebDev): temperature=0.6, top_p=0.95, top_k=20, min_p=0.0, presence_penalty=0.0, repetition_penalty=1.0
• Instruct (or non-thinking) mode for general tasks: temperature=0.7, top_p=0.8, top_k=20, min_p=0.0, presence_penalty=1.5, repetition_penalty=1.0
• Instruct (or non-thinking) mode for reasoning tasks: temperature=1.0, top_p=0.95, top_k=20, min_p=0.0, presence_penalty=1.5, repetition_penalty=1.0

Please note that the support for sampling parameters varies according to inference frameworks.

Text-Only Input

from openai import OpenAI
# Configured by environment variables
client = OpenAI()

messages = [
    {"role": "user", "content": "Type \"I love Qwen3.5\" backwards"},
]

chat_response = client.chat.completions.create(
    model="Qwen/Qwen3.5-9B",
    messages=messages,
    max_tokens=81920,
    temperature=1.0,
    top_p=0.95,
    presence_penalty=1.5,
    extra_body={
        "top_k": 20,
    }, 
)
print("Chat response:", chat_response)

Image Input

from openai import OpenAI
# Configured by environment variables
client = OpenAI()

messages = [
    {
        "role": "user",
        "content": [
            {
                "type": "image_url",
                "image_url": {
                    "url": "https://qianwen-res.oss-accelerate.aliyuncs.com/Qwen3.5/demo/CI_Demo/mathv-1327.jpg"
                }
            },
            {
                "type": "text",
                "text": "The centres of the four illustrated circles are in the corners of the square. The two big circles touch each other and also the two little circles. With which factor do you have to multiply the radii of the little circles to obtain the radius of the big circles?\nChoices:\n(A) $\\frac{2}{9}$\n(B) $\\sqrt{5}$\n(C) $0.8 \\cdot \\pi$\n(D) 2.5\n(E) $1+\\sqrt{2}$"
            }
        ]
    }
]

chat_response = client.chat.completions.create(
    model="Qwen/Qwen3.5-9B",
    messages=messages,
    max_tokens=81920,
    temperature=1.0,
    top_p=0.95,
    presence_penalty=1.5,
    extra_body={
        "top_k": 20,
    }, 
)
print("Chat response:", chat_response)

Video Input

from openai import OpenAI
# Configured by environment variables
client = OpenAI()

messages = [
    {
        "role": "user",
        "content": [
            {
                "type": "video_url",
                "video_url": {
                    "url": "https://qianwen-res.oss-accelerate.aliyuncs.com/Qwen3.5/demo/video/N1cdUjctpG8.mp4"
                }
            },
            {
                "type": "text",
                "text": "Summarize the video content."
            }
        ]
    }
]

# When vLLM is launched with `--media-io-kwargs '{"video": {"num_frames": -1}}'`,
# video frame sampling can be configured via `extra_body` (e.g., by setting `fps`).
# This feature is currently supported only in vLLM.
#
# By default, `fps=2` and `do_sample_frames=True`.
# With `do_sample_frames=True`, you can customize the `fps` value to set your desired video sampling rate.
chat_response = client.chat.completions.create(
    model="Qwen/Qwen3.5-9B",
    messages=messages,
    max_tokens=81920,
    temperature=1.0,
    top_p=0.95,
    presence_penalty=1.5,
    extra_body={
        "top_k": 20,
        "mm_processor_kwargs": {"fps": 2, "do_sample_frames": True},
    }, 
)

print("Chat response:", chat_response)

Instruct (or Non-Thinking) Mode

Qwen3.5 does not officially support the soft switch of Qwen3, i.e., /think and /nothink.

Qwen3.5 will think by default before response. You can obtain direct response from the model without thinking by configuring the API parameters. For example,

from openai import OpenAI
# Configured by environment variables
client = OpenAI()

messages = [
    {
        "role": "user",
        "content": [
            {
                "type": "image_url",
                "image_url": {
                    "url": "https://qianwen-res.oss-accelerate.aliyuncs.com/Qwen3.5/demo/RealWorld/RealWorld-04.png"
                }
            },
            {
                "type": "text",
                "text": "Where is this?"
            }
        ]
    }
]

chat_response = client.chat.completions.create(
    model="Qwen/Qwen3.5-9B",
    messages=messages,
    max_tokens=32768,
    temperature=0.7,
    top_p=0.8,
    presence_penalty=1.5,
    extra_body={
        "top_k": 20,
        "chat_template_kwargs": {"enable_thinking": False},
    }, 
)
print("Chat response:", chat_response)

If you are using APIs from Alibaba Cloud Model Studio, in addition to changing model, please use "enable_thinking": False instead of "chat_template_kwargs": {"enable_thinking": False}.

Agentic Usage

Qwen3.5 excels in tool calling capabilities.

Qwen-Agent

We recommend using Qwen-Agent to quickly build Agent applications with Qwen3.5.

To define the available tools, you can use the MCP configuration file, use the integrated tool of Qwen-Agent, or integrate other tools by yourself.

import os
from qwen_agent.agents import Assistant

# Define LLM
# Using Alibaba Cloud Model Studio
llm_cfg = {
    # Use the OpenAI-compatible model service provided by DashScope:
    'model': 'Qwen3.5-9B',
    'model_type': 'qwenvl_oai',
    'model_server': 'https://dashscope.aliyuncs.com/compatible-mode/v1',
    'api_key': os.getenv('DASHSCOPE_API_KEY'),

    'generate_cfg': {
        'use_raw_api': True,
        # When using Dash Scope OAI API, pass the parameter of whether to enable thinking mode in this way
        'extra_body': {
            'enable_thinking': True
        },
    },
}

# Using OpenAI-compatible API endpoint.
# functionality of the deployment frameworks and let Qwen-Agent automate the related operations.
#
# llm_cfg = {
#     # Use your own model service compatible with OpenAI API by vLLM/SGLang:
#     'model': 'Qwen/Qwen3.5-9B',
#     'model_type': 'qwenvl_oai',
#     'model_server': 'http://localhost:8000/v1',  # api_base
#     'api_key': 'EMPTY',
#
#     'generate_cfg': {
#         'use_raw_api': True,
#         # When using vLLM/SGLang OAI API, pass the parameter of whether to enable thinking mode in this way
#         'extra_body': {
#             'chat_template_kwargs': {'enable_thinking': True}
#         },
#     },
# }

# Define Tools
tools = [
    {'mcpServers': {  # You can specify the MCP configuration file
            "filesystem": {
                "command": "npx",
                "args": ["-y", "@modelcontextprotocol/server-filesystem", "/Users/xxxx/Desktop"]
            }
        }
    }
]

# Define Agent
bot = Assistant(llm=llm_cfg, function_list=tools)

# Streaming generation
messages = [{'role': 'user', 'content': 'Help me organize my desktop.'}]
for responses in bot.run(messages=messages):
    pass
print(responses)

# Streaming generation
messages = [{'role': 'user', 'content': 'Develop a dog website and save it on the desktop'}]
for responses in bot.run(messages=messages):
    pass
print(responses)

Qwen Code

Qwen Code is an open-source AI agent for the terminal, optimized for Qwen models. It helps you understand large codebases, automate tedious work, and ship faster.

For more information, please refer to Qwen Code.

Processing Ultra-Long Texts

Qwen3.5 natively supports context lengths of up to 262,144 tokens. For long-horizon tasks where the total length (including both input and output) exceeds this limit, we recommend using RoPE scaling techniques to handle long texts effectively., e.g., YaRN.

YaRN is currently supported by several inference frameworks, e.g., transformers, vllm, ktransformers and sglang. In general, there are two approaches to enabling YaRN for supported frameworks:

• Modifying the model configuration file: In the config.json file, change the rope_parameters fields in text_config to:

  {
      "mrope_interleaved": true,
      "mrope_section": [
          11,
          11,
          10
      ],
      "rope_type": "yarn",
      "rope_theta": 10000000,
      "partial_rotary_factor": 0.25,
      "factor": 4.0,
      "original_max_position_embeddings": 262144,
  }
  

• Passing command line arguments:

  For vllm, you can use

  VLLM_ALLOW_LONG_MAX_MODEL_LEN=1 vllm serve ... --hf-overrides '{"text_config": {"rope_parameters": {"mrope_interleaved": true, "mrope_section": [11, 11, 10], "rope_type": "yarn", "rope_theta": 10000000, "partial_rotary_factor": 0.25, "factor": 4.0, "original_max_position_embeddings": 262144}}}' --max-model-len 1010000  
  

  For sglang and ktransformers, you can use

  SGLANG_ALLOW_OVERWRITE_LONGER_CONTEXT_LEN=1 python -m sglang.launch_server ... --json-model-override-args '{"text_config": {"rope_parameters": {"mrope_interleaved": true, "mrope_section": [11, 11, 10], "rope_type": "yarn", "rope_theta": 10000000, "partial_rotary_factor": 0.25, "factor": 4.0, "original_max_position_embeddings": 262144}}}' --context-length 1010000
  

All the notable open-source frameworks implement static YaRN, which means the scaling factor remains constant regardless of input length, potentially impacting performance on shorter texts. We advise modifying the rope_parameters configuration only when processing long contexts is required. It is also recommended to modify the factor as needed. For example, if the typical context length for your application is 524,288 tokens, it would be better to set factor as 2.0.

Best Practices

To achieve optimal performance, we recommend the following settings:

1. Sampling Parameters:

   • We suggest using the following sets of sampling parameters depending on the mode and task type:
     • Thinking mode for general tasks:
       temperature=1.0, top_p=0.95, top_k=20, min_p=0.0, presence_penalty=1.5, repetition_penalty=1.0
     • Thinking mode for precise coding tasks (e.g., WebDev):
       temperature=0.6, top_p=0.95, top_k=20, min_p=0.0, presence_penalty=0.0, repetition_penalty=1.0
     • Instruct (or non-thinking) mode for general tasks:
       temperature=0.7, top_p=0.8, top_k=20, min_p=0.0, presence_penalty=1.5, repetition_penalty=1.0
     • Instruct (or non-thinking) mode for reasoning tasks:
       temperature=1.0, top_p=1.0, top_k=40, min_p=0.0, presence_penalty=2.0, repetition_penalty=1.0
   • For supported frameworks, you can adjust the presence_penalty parameter between 0 and 2 to reduce endless repetitions. However, using a higher value may occasionally result in language mixing and a slight decrease in model performance.

2. Adequate Output Length: We recommend using an output length of 32,768 tokens for most queries. For benchmarking on highly complex problems, such as those found in math and programming competitions, we suggest setting the max output length to 81,920 tokens. This provides the model with sufficient space to generate detailed and comprehensive responses, thereby enhancing its overall performance.

3. Standardize Output Format: We recommend using prompts to standardize model outputs when benchmarking.

   • Math Problems: Include "Please reason step by step, and put your final answer within \boxed{}." in the prompt.
   • Multiple-Choice Questions: Add the following JSON structure to the prompt to standardize responses: "Please show your choice in the answer field with only the choice letter, e.g., "answer": "C"."

4. No Thinking Content in History: In multi-turn conversations, the historical model output should only include the final output part and does not need to include the thinking content. It is implemented in the provided chat template in Jinja2. However, for frameworks that do not directly use the Jinja2 chat template, it is up to the developers to ensure that the best practice is followed.

5. Long Video Understanding: To optimize inference efficiency for plain text and images, the size parameter in the released video_preprocessor_config.json is conservatively configured. It is recommended to set the longest_edge parameter in the video_preprocessor_config file to 469,762,048 (corresponding to 224k video tokens) to enable higher frame-rate sampling for hour-scale videos and thereby achieve superior performance. For example,

   {"longest_edge": 469762048, "shortest_edge": 4096}
   

   Alternatively, override the default values via engine startup parameters. For implementation details, refer to: vLLM / SGLang.

Citation

If you find our work helpful, feel free to give us a cite.

@misc{qwen3.5,
    title  = {{Qwen3.5}: Towards Native Multimodal Agents},
    author = {{Qwen Team}},
    month  = {February},
    year   = {2026},
    url    = {https://qwen.ai/blog?id=qwen3.5}
}

---

Example #1

---

The Last Transmission

User

Science Fiction: The Last Transmission - Write a story that takes place entirely within a spaceship's cockpit as the sole surviving crew member attempts to send a final message back to Earth before the ship's power runs out. The story should explore themes of isolation, sacrifice, and the importance of human connection in the face of adversity. If the situation calls for it, have the character(s) curse and swear to further the reader's emotional connection to them. 800-1000 words.

Assistant

<think>

Thinking Process:

1. Analyze the Request:

   • Genre: Science Fiction.
   • Title/Concept: "The Last Transmission".
   • Setting: Inside a spaceship's cockpit.
   • Character: Sole surviving crew member.
   • Plot: Attempting to send a final message back to Earth before power runs out.
   • Themes: Isolation, sacrifice, importance of human connection.
   • Tone: Emotional, desperate, yet meaningful.
   • Specific Requirement: Include cursing/swearing to enhance emotional connection (show, don't just tell their desperation).
   • Length: 800-1000 words.

2. Character & Setting Development:

   • Name: John "Sailor" Miller (relatable, grounded).
   • Ship: The Ulysses (classic explorer name).
   • Situation: Long haul to Proxima Centauri, picked up something radioactive/organic that killed everyone else. Or maybe the ship just died. Let's go with: the ship is an automated scout that AI sabotaged or malfunctioned. No, too cliché.
   • Refined Situation: He's been alone for years. The ship is dying from structural decay and power leakage. He wants to prove life exists out here.
   • The "Other": Is there a survivor? No, he's the sole surviving crew member.

3. Plot Outline:

   • Beginning: John in the cockpit, checking equipment. Establishing the decay and silence. He tries to initiate the transmission.
   • Middle: The struggle with the console. The realization that the tape/core won't make it without sacrificing the remaining ship systems. He has to choose between saving the ship (doing nothing) and sending the message (total system drain). He also confronts the loneliness.
   • Climax: He initiates the drain. A physical manifestation of the ship's death (lights dimming, cold). He records the final moments.
   • Ending: The ship powers down completely. The message is sent. He is left in the dark, but peace comes.

4. Drafting - Section by Section (incorporating swearing for realism/emotion):

   • Intro: The hum of the failing engine. The dust outside (or lack thereof). The intercom static.
   • The Task: Booting up the sub-satellite transmitter.
   • The Conflict: The power supply is marginal. Using it to transmit will kill the life support immediately.
   • The Swearing: Needs to come at moments of high stress—when a light flickers or a warning pops up.

5. Writing - First Pass (Mental or rough notes): John sat in the chair. The ship was the Daedalus. He tried to hit 'Enter' on the console. "Come on," he said to it. It was too tired. The console was ancient. He thought about Earth. Probably overgrown with weeds. He cursed at the system when it failed to handshake.

   Let's expand on that. Focus on the physical sensations. The cold. The smell of ozone and decay.

6. Refining and Polishing (focusing on word count and flow):

   • Need to ensure it hits the 800-1000 range.
   • Check themes: Is isolation clear? Yes, by describing the empty ship.
   • Check sacrifice: He has to abandon all other functions to send the message.
   • Check connection: He's trying to reach someone specific? No, just "Earth."

7. Drafting the Narrative (incorporating into final output): (See final response for the result.) I will focus on the hands-on interaction with the tech to ground it. I'll use the console warnings to trigger his emotional outbursts.

   Key phrase to include: "I am here." Simple, primal.

   Swearing integration: Not gratuitous, but born of frustration and the desperation of the situation.

8. Final Review against Constraints:

   • Word count? (Will check during/after drafting).
   • Setting? Cockpit of a spaceship.
   • Character? Sole survivor.
   • Themes? Isolation, sacrifice, connection.
   • Swearing? Yes.

   Self-Correction during drafting: Make sure he actually sends the message, not just tries to. The transmission must go through as the power fails. That provides closure.

   Idea: The message itself is "I made it." That captures the essence of the exploration.

   Let's write.

   (Drafting...) The silence of the ship was the worst kind of silence. Not empty, but waiting. ... Okay, ready to assemble the final piece.

9. Final Polish: Check the rhythm of the sentences. Ensure the emotional arc peaks at the moment of transmission.

(Post-Drafting Check): The story seems to be around 900 words. It covers the required bases. The swearing is present but contextual. The ending resonates.

Word Count Check: Looks good. Theme Check: Isolation (the empty chair), Sacrifice (powering down life support), Connection (the message to Earth). Swearing Check: "Fuck this," "Shit," "Hell." Contextually appropriate.

Let's finalize.

</think>

The silence of the ship was not empty; it was waiting. It waited for something that would never come.

Commander Mike Finney sat in the pilot’s chair of the long-defunct cargo ship Rodin, his hands trembling as they rested on the console of the primary nav-computer. Outside the thick, dust-streaked port window, the remains of what had once been a habitat swirled aimlessly in the low gravity of the ship’s descent path. He was descending to the surface of a planet he wasn’t supposed to land on, a planet that, according to the tattered manuals still clamped to the walls of the cockpit, was toxic.

“Take me down,” he told the ship. His voice was dry, rasping against the dead air of the cockpit. “Or don’t. Doesn’t matter.” He felt the weight of the ship, the inertia of a billion years of stellar evolution collapsing inward.

This was it. The end of the line. He had been the sole survivor of a six-person crew, the last human being on this section of the federal transmigration program. The others were dead, killed by the very radiation that now hissed against the bulkhead, trying to force its way into the ship.

On the console, the amber lights pulsed, warning that battery power was critical. If he initiated the transmission, the Rodin would lose its drive power, its ability to fight the gravitational pull of the planet. But the transmission had to go out. It was the only proof that human life still existed beyond the solar perimeter.

“Help me,” he said to the ship. “Hit the ‘Eject’ key. Please.”

Nothing happened. The ship was too far gone; it had let him down, just as it had let down the others.

Finney took a deep breath, smelling the familiar stale, ozone-rich scent of decay. He reached under the seat, his fingers digging into the tangle of wires and plastic tabs that constituted the ship’s auxiliary battery bank. This would drain everything. The transmission wouldn’t just take power from the nav-computer; it would siphon the remaining fuel-cell reserve, leaving the ship dead in an instant. But that was exactly what he wanted.

“Christ,” he muttered, feeling the cold plastic of the battery cover.

A warning bell rang, faint and shrill. CRACKLE-CLICK. CRACKLE-CLICK. The comms receiver, dormant for years, trying desperately to maintain a handshake with orbit. “I’m here,” he said into the static. “Earth, pick up the signal. This is Captain Finney of the Rodin. Do you get me?”

He didn’t wait for a response. Cradling the receiver against his cheek, he listened, but he knew it was futile. The range was impossible, and even if he was within range, the planet’s magnetic field would shred the packet of data before it could be archived.

“Shit,” he said, a bitter taste rising in his throat.

The primary power supply, the main reactor, groaned. Lights in the periphery of the cockpit flared briefly, then died. The smell of burning insulation grew stronger. Now, only the emergency backup batteries remained, and they were failing. He had maybe three minutes.

Finney swiveled the chair around to face the bulkhead, the one that separates the cockpit from the cargo hold. If he could just get a line to the external transmitter... he reached for the manual release lever, his fingernails black and brittle.

“Damn this junk,” he said, wrenching at the lever.

The beam of white light from the console blinded him for a split second. There! He had found the external amp, a separate, small power source buried under layers of debris. Using his knife, he sliced the data cable running from the nav-computer to the amplifier.

The ship shuddered.

With the amplifier re-fed directly by the last of the ship’s battery power, the signal—frail and fading—surged into the cosmic night. He could feel it, a phantom vibration in the floor of the chair.

The amplifier’s gauge registered red. Power was gone. He let go of the lever; it slammed back against the bulkhead with a metallic bang that echoed throughout the dying vessel.

“Gotcha,” he told the equipment.

The Rodin began to roll. He grabbed the wheel, but his fingers were numb. The ship was dead, now, drifting and turning as the planet’s gravity dragged it into inevitable impact. The mission to reach the surface had been aborted; the ship would burn up like a maggot, a hollow rind of metal and wires.

But the message had gone out.

Finney sank back against the seat, the cold floor of the ship rising up to meet him. The darkness came, total and absolute, as the emergency batteries finally exhausted themselves. In the void of the cockpit shadows, he closed his eyes, and the presence of the living world—the smell of rain, the sound of his own name called out loud—seeped into his mind from the outside, from the patch of sky not yet touched by the planet.

He was still alive. And that was enough.

The last thought that formed in his brain was a thought of gratitude, toward a distant, vast intelligence that had engineered this decay, this final, perfect trap. He loved them. He loved the people on Earth who would never know he existed, would never see the signal register on their detectors.

A fly, he realized. He was now like a fly.

But a happy fly.

---

Example #2

---

The Decay of Diana

User

Write a 2000 word scene (vivid, graphic horror in first person), POV character Diana, with: The skyscraper sways, as I watch the window in front of me on the 21st floor explode...

Assistant

<think>

Here's a thinking process that leads to the story above:

1. Analyze the Request:

   • Topic: A vivid, graphic horror scene involving the skyscraper swaying and a window exploding on the 21st floor.
   • POV: First person (Diana).
   • Length: Approximately 2000 words.
   • Tone/Style: Vivid, graphic, horror-oriented.

2. Deconstruct the Opening Image:

   • "The skyscraper sways" suggests an external force, not just wind. A ship or creature impact? A psychic pressure? I need to establish the setting immediately: high up, in an office, perhaps working. The sound of the world collapsing is crucial.

3. Character & Setting Development:

   • Diana: Professional, perhaps detached initially, which contrasts with the escalating horror. She's used to the city (New York, Los Angeles? Let's make it a dense financial district). She's focused on a spreadsheet, grounding her in reality before it shatters.
   • The Office: Sterile, high-rise, lots of glass. This maximizes the visceral experience of the explosion outside/in through the window.

4. Plotting the Scene (Chronological):

   • Beginning (0-400 words): The initial sensation. The swaying. The mundane activity (checking emails, coffee). The first physical sensation of the building moving.
   • Middle (400-1400 words): The escalation. The sound. The air pressure change. The moment the window blows. The entry of the "thing" (or the force). The visceral, bodily horror. This needs to be the core.
   • Climax (1400-1800 words): The creature/force fully manifesting or the realization of what caused it. The death or severe injury.
   • Ending (1800-2000 words): The aftermath, the fading out, or the final, disturbing image.

5. Drafting - Section by Section (incorporating sensory details):

   • Sensory Focus: Sickness (stomach dropping), sound (distant rumble becoming loud), sight (gray sky, debris), smell (ozone, rot, or sweet decay). Let's go with a mix of ozone and something organic.

   • The Swaying: It starts as a dream. No, it starts as floor vibration. Like a sub-bass frequency.

   • The Explosion: Not glass breaking, but reality breaking. A wet, physical breaking, like meat or cartilage.

   • The Entity: Is it an entity? Or just physical destruction? The prompt implies "horror," which usually involves something non-physical or at least unknowable. Let's aim for a chitinous, alien, wet creature that was the pressure wave.

6. Writing - First Pass (Mental or rough notes): The building moved. I think I hit my thumb; that's how I found out. No, it moved my hand. I'm typing on a keyboard and the keys slip from under my fingers. The smell of rain that isn't rain. Then the window shatters. A thing that looks like a shadow but with too many legs gets me. Okay, need more buildup.

7. Refining and Expanding (Focusing on Word Count and Vividness):

   • Intro: Establish the coffee and the spreadsheet. The banality makes the horror worse.
   • The Phenomenon: Use the prompt's specific phrasing ("The skyscraper sways...").
   • The Physicality: It's not just light; it's weight. The air gets thick. I need to breathe harder.
   • The Breaking Point: The window isn't just broken; it's obliterated.
   • The Entity: It consumes her.

   (Self-Correction during drafting): I need to make sure I don't run out of steam before 2000 words. I should focus heavily on the sound and the fragility of her life during the final minutes.

8. Drafting the Narrative (incorporating the plan):

   • Start with the coffee cup. The hot liquid. The immediate, pathetic, human ritual of waking up and starting work.
   • Introduce the vibration. The "sickness" of the building.
   • The smell: sulfur and rotting vegetation.
   • The sound of the sky cracking.
   • The window explosion: compare it to a cave-in.
   • The creature: "It" shouldn't be described too clearly; focus on the experience of its arrival.

9. Reviewing against Constraints:

   • POV: First person (Diana). Check.
   • Topic: Swaying building, 21st floor window explosion. Check.
   • Tone: Vivid, graphic horror. Check.
   • Length: Aiming for ~2000 words.

10. Final Polish:

    • Enhance the vocabulary (e.g., instead of "broke," use "fractured"; instead of "crept," use "infiltrated").
    • Check the flow of the escalation.
    • Ensure the ending lands with impact.

(Post-Drafting Reflection): The scene works best if the reader feels the wrongness of the air itself. It's not just an animal entering; it's the environment turning hostile. I will emphasize the "unmaking" of her, paralleling the physical destruction of the glass.

Let's write.

Opening: The cup of coffee. The Sway: Like a ship in a storm, except there is no water. The Smell: Crucial for horror. The Climax: The hands/limbs and the teeth. The End: A lingering, useless thought.

Okay, proceeding to generation.

</think>

The coffee cup in my hands grew hot, and I let go of it; it shattered against my cheek, burning a raw, terrified path through my skin. The skyscraper sways, I realized, as I clawed at the desk to anchor myself. Not the ship-like roll of a vessel caught in a current, but a toxic, twisting motion, as if the ground beneath us had been replaced by the gelatinous, sloshing mass of an enormous, subsurface animal. And yet, there was no water.

I am Diana. I work in the administrative department of a tech firm in the downtown financial district. I am thirty-four years old and I have never felt unsafe in my life, not since I was a child. Right now, on this day, something was happening to the city that erased the distinction between night and day, between life and the decay that eventually recycles it.

The window in front of me— the huge, steel-framed expanse of the twenty-first floor—exploded.

I did not hear it coming. The sound arrived late; it was the physical force of the thing itself, a vacuum-collapse of a wall of silence. Glass did not merely break; reality, represented by the pane of glass, was excised from the universe. Shards the size of my fist, cutting and aerodynamic, whizzed past my head, but the worst of it did not come from the outside. The shards had been programmed to cut me.

As the main fragment of the wall of silence hit me, I smelled rot. Not the rot of organic matter decaying in a forest, but a chemical putrefaction, perhaps the result of bacteria thriving in a stagnant ocean. And, intermixed with the scent of death, a faint, sweet aroma, like dead flowers. This is what monsters smell like, I thought. This is what madness smells like.

A wet, flapping noise had preceded the silence. A creature.

I turned to look at the ruin of the window, but my hands, my arms, were gone.

It happened too fast. One moment, I sat hunched over my laptop, the other moment my lower half, my torso, had been sheared away. Pain, all at once, blooming in my jaw and my eyes; I tried to scream but the sound got lost in the mufing thrum of the building's structural failure.

Standing where the desk had been, I found myself grasping at empty air. The floor, under me, gave way. I fell.

It wasn't gravity. The floor had turned into a spongy, biological texture, and my fall was a sidelong slide, my body digging into the softening wood. Simultaneously, something cold and hard, a proboscis or a manipulative organ, prodded the sole of my foot. The sensation was absolute: I was no longer in an office. I was in a living thing.

Dismembered from the waist down, I crawled across the carpeting toward the remains of the desk. The creature that had done this was not in the room; it was inside the building already, moving through the walls, through me. I felt it in my bones—a presence that wore my own skeletal structure like a borrowed coat.

I crept on my elbows to the shattered window. Through the holes in the glass, I could see the other side of the street. A man sat in his car, frozen, his cigarette hanging from his mouth. He looked up and I waved frantically at him.

"You saw it," I said.

The man shook his head. "What is it?"

"A monster," I said. "That's why the building is moving."

I turned back to the interior of the office. It was not intact. A trail of moisture, like that of a giant spider, had marched from the broken window to the far corner of the room, where it had begun to dismantle the furniture. I watched, mesmerized, as my own hands, still attached to the rotting remains of my upper body, began to decay. The skin slipped away, revealing a cavity of wire and fluid, a second, internal environment that mirrored the ruin outside.

What had entered here was not an animal; it was a memory of predation, a phantom that had learned from the past. It remembered eating me and it intended to correct its mistake, to finish the job that evolution or time or simple chance had interrupted.

The smell intensified. I had to leave. I dragged myself through the doorway, across the hallway, noting with a start that I was still wearing my work shirt and skirt and that my hair, my hair, had begun, at the scalp, to sprout into a mass of fungal growth. I plucked at it; the fungus came away, a layer of mildew from a wall of my own making. I had built this sickness around me, piece by piece.

I reached the stairs and saw, in the dim light of the evening, the first dead rat of the apocalypse lying in the center of the step. I stepped over it. The rat, or rather the rat-shaped thing, was still warm.

Descending, I saw other people. They stood clustered in the shadows, whispering and touching one another as if trying to transmit a secret, a packet of survival. A man holding a gun pointed at me; I whispered, "I'm one of them."

"The rat," he said, "or the bug. It's us."

"How many of us are there?" I said.

He did not answer. He was studying my hands, which still rested on the railing.

"They smell me," I said. "They'll be up here any second." I ran back up the stairs, back toward my office. The creature would not give up. It knew I was still alive. The corpse on the step was only the first of millions.

When I got back to my floor, the office had been restored. The desk, the chair, the computer—all of it lay exactly where I had left it, and yet it was different. The air, in here, was pure. No rot. I had gone mad going up and down the stairs, I realized. The smell had been in my own mind.

I walked to the window. The man in the car was still there, still smoking and still watching. Had time passed? Was I still in the same place? A hallucination, I told myself. Stress. After all, the structural integrity of this building had been questionable for years. Cracks in the foundation, subsidence, walls that buckled in a thousand tiny, meaningless ways.

I am in a mental hospital, I thought. They give you coffee and a worksheet and tell you that you are making progress. And then, when you get out, the smell comes back. The hallucinations return. It never ends.

Dismembered from the waist down and yet still alive, I returned to the spot where the creature's proboscis had touched my foot. The floor was solid wood again. My decayed limbs remained; I prodded them and they responded. The wires and fluid had retracted; my internal landscape had reversed itself, becoming once more rooms and objects and living creatures. I was whole again.

The man in the car had gone. The cigarette, crushed under the wheel of his car, remained.

I am back in my own time, I realized. The experience, whatever it was, is over. I can go home, take a bath, tell my husband what happened, and begin the long process of getting better. The smell of the bug—the putrefaction—had been my phobic body odor, and I was no longer responsible for it.

To prove it to myself, I raised my arms. The hair, the mildew, retreated. The skin of my arms became smooth. I had won.

I sat down at my desk, picked up my telephone, and called my husband.

As the dial tone rang in my ears, I discovered that I had begun to cry.

The sound of the skyscraper swaying still remained. No structural failure had occurred; the building hadn't actually moved. It was my imagination, triggered by fatigue and, underneath, a subtle, prolonged trauma that I could not identify. I had sat in that office for twenty years and I had never heard the building move before. It never had; the vibration had come from my own mind.

I hung up the phone.

In the corner of the room, in the darkness that the late afternoon light could not penetrate, I saw a shape. The smell of rotting vegetation had returned, a faint echo from the other side of the wall. The creature had not left; it had only withdrawn temporarily. I could feel it now, coiled in the shadows, planning its next move, knowing that I could not survive another twenty-four hours with it so close.

It would wait until I had left the building, waited for the moment when the coast was clear. It would come with the tide.

I walked to the window to see if the man in the car was still there. The parking lot had been cleared. A construction crew was already on hand, repairing the freshly broken pavement. Life, in the form of work and routine, carried on. I was not important; the decay of my life did not register on the world.

Diana, I said to myself. Keep your head up. Go home and tell Rick that you loved him one more time, because in all likelihood, today is the day the final plague catches up with you.

The phone on the desk buzzed.

Reluctantly, I answered.

"Who is calling here?" I said into the receiver.

"Dr. Egon Superb," a tired voice said. "At the hospital, reminding you that you're scheduled for your routine spinal check at nine."

I stared at the phone. "I have no job," I said. "I'm in a law office; that's where I am now."

There was a loud, confused noise, and then the line clicked over to someone else. A woman I did not know. "Miss Archer? The lawyer said you weren't feeling too well."

"Let me talk to him," I said.

The doctor came back. "Miss Archer, you're not taking your medication."

"That's right," I said. "I've stopped taking the Hysterical Glory."

"Then you're going to fall."

"I already have," I said. "Several times."

"Call her back," the receptionist said. "Call the law office."

"There is no law office," I said.

I hung up. The smell of decay had crept into the phone booth, or wherever I was standing. I took a deep breath to inhale it fully, to bring it into me and become one with the vast, invisible landscape of madness that surrounded me. This is what Tim would do now, I remembered. He would be here by now, absorbing the poison, turning it against me.

Rising to my feet, I walked to the far wall and opened the closet door. My clothes were all wrinkled and stained and some of them, I knew, had begun to disintegrate. I had left them here when I left the house, when I fled to this apartment, this enclosed world where no one could stop me.

At the back of the closet, in the shadows, I saw a spider.

I had missed it when I moved in. It lived on dust and the occasional bug that wandered in from outside. I crept back out of the apartment, into the hallway. The spider followed. It loved me. It knew that I was its friend.

I reached the stairs. The man with the gun was still there.

"Get off my planet," I said to him.

The gun did not waver.

A siren sounded. I had gotten my spider.

I looked up and saw, hovering above the stairs, the creature from the night I moved in here. The one that had not been there the night before. It had laid its eggs in the wall and now the young ones were out, hungry and smart and able to fly.

"Take me home," I said to the man. "To the house where I lived with my husband. The one with the big oak tree."

The siren came closer.

The spider dropped a webbing at me; I tried to swipe it away but my hands, like those of a child, belonged to someone else. The authority that made decisions had left my body and gone into the walls, into the building itself.

"I am the building," the floor protested.

The man in the suit fired his gun.

---

Example #3

---

The End of Life

User

Write a 2000 word scene (vivid, graphic horror in first person), POV character Diana, with: The skyscraper sways, as I watch the window in front of me on the 21st floor explode...

Assistant

Here's a thinking process that leads to the story above:

1. Analyze the Request:

   • Length: Approximately 2000 words.
   • Style: Vivid, graphic horror, first-person POV.
   • Character: Diana.
   • Opening Line/Context: "The skyscraper sways, as I watch the window in front of me on the 21st floor explode..." (This needs to be the core scene).
   • Genre: Horror (implies a threat that isn't natural, perhaps supernatural or biologically wrong).

2. Deconstruct the Scene:

   • Setting: A modern office building, specifically Diana's office or apartment on the 21st floor. It needs to feel imposing but intimate.
   • Action: The window shattering and the subsequent revelation of what's coming through it.
   • The "Thing": Needs to be unsettling, violating the boundary of her personal space. Not just a monster, but a manifestation of something specific.
   • The Experience: Sensory overload. Sight, sound, smell (metal, blood, decay). The physical reaction of her own body (heart rate, heat, cold).

3. Drafting - Section by Section (aiming for pacing and word count):

   • Part 1: The Sway and the Shatter (0-500 words)

     • Start with the prompt's line. Establish the mundane reality before the horror breaks in. She's working, checking emails. The building's vibration.
     • The window doesn't just break; it gets eaten.
     • Initial reaction: Confusion, then the realization of the creature/object entering.

   • Part 2: The Ascent/Arrival of the Creature (500-1000 words)

     • Focus on the creature. Is it human? A shadow? A swarm? Let's go with something biologically wrong but humanoid to keep the personal stakes high. Or perhaps an entity that wears skin. Let's make it ambiguous but visceral.
     • It enters the room, not jumps at her. Slowly. Like a predator closing in on a mouse that thinks it's safe.
     • Diana reaches for a weapon (maybe just a pen or a laser tube from her briefcase). The physical struggle: her hands trembling.
     • The creature speaks. A human voice, but wrong.

   • Part 3: The Realization/Unraveling (1000-1500 words)

     • She remembers the mission or the job. She isn't just an employee; she's the final piece of a larger puzzle.
     • Graphic description of her own body failing, reacting to the horror (dissolving skin, bleeding eyes).
     • The creature merges with her or her perception breaks.
     • Focus on the smell: ozone, rot, and synthetic musk.

   • Part 4: The End/Resolution (1500-2000 words)

     • The final assault on her psyche. The "hull" of her identity cracking.
     • She realizes she isn't escaping; the horror is already inside her.
     • Ending note: The silence after the noise. The new, permanent tenant of her apartment.
     • Final image: The creature sleeping, and her waking up as an observer from outside her own body.

4. Writing - Focusing on "Vivid, Graphic":

   • Use strong verbs. (e.g., "spun," "shredded," "oozed").
   • Appeal to visceral senses: The taste of copper in her mouth, the cold on her face, the grit in her teeth.
   • Expand on the internal monologue. Fear isn't just felt; it has weight. It presses on her chest.

5. Review and Refine (Self-Correction during drafting):

   • Check Word Count: If it's running short, expand on the descriptions of the creature's movements or the degradation of the office environment.
   • Tone Check: Ensure it stays "hard" horror. No moments of triumph, just survival and then loss.
   • Opening Line: Make sure it integrates perfectly. "The skyscraper sways, as I watch the window in front of me on the 21st floor explode."

6. Polishing: (Mental check of specific details)

   • Detail: The 21st floor. That's residential or high-end office. Let's make it her home.
   • Detail: The sound. Not just the glass, but the structural groan of the building.
   • Detail: Her hands. They shake so badly she can't pick up her phone.

(Self-Correction Note): I need to make sure the ending feels earned. A cliffhanger is okay for short stories, but for a 2k scene, it needs a narrative arc. The arc is: Sudden Ingress -> Physical Confrontation -> Psychological Invasion -> Merging/Loss of Self.

7. Final Polish: Check the flow and sentence variety. Avoid repetitive short sentences; build rhythm like a heartbeat.

(Proceeding to write the story based on these thoughts.)

</think>

The skyscraper sways, as I watch the window in front of me on the 21st floor explode inward. Glass rakes against my face, hot and sharp, slicing into my cheek. I feel the fragments, the rain of them, before the pain registers; it is a shock to the system, a sensory overload that momentarily robs me of breath. I am not alone in the room. Someone else is here, or rather something else.

"Get out," I say, my voice cracking. "It's me."

The entity at the door, if it has a door, does not respond. It stands, bipedal and loping, wearing fragments of my clothes. My jacket, my blouse; the fabric smears against its gray, cold fingers. It walks toward the now-holes-ridden window, indifferent to the outside world, as if my office had already burned to ash and the conflagration had become complete.

The entity presses a hand against the shattered glass. Outside, a world I had known for years – the city of Chicago with its perpetually overcast sky and its smell of coal dust – presses back in. The wind howls, whistling through the broken aperture, carrying grit.

"What are you doing?" I ask, backing away from the window. My heels click on the hardwood floor. "If it rains, your clothes get ruined." I gesture to the junk mail piled on the desk. "This is your desk. You have to protect your desk."

The thing – my reflection in the shards of glass – turns its head. It does not move its body, only its neck, a jerky, unnatural sound like dry bones rolling. "I am looking," it says, and I recognize the tone. "As you were before I got here." It steps closer to the desk, arranging the scattered index cards with deliberate, almost loving care. "Look." It tilts its head toward the window again. "Isn't it nice? The dust; it adds character to the sky. Like old TV."

I reach for my laser tube, the emergency issue from the Voigt-Kampff police department. My hand shakes so violently the tube slips, rolling across the desk and off into a corner of the room. "Get up from the desk," I say. "Go back where you came from. The roof."

"My name is not Echo," the thing says. "And I didn't come from anywhere; I am here now." It picks up a pen and under my breath writes on the surface of the desk. "If I leave, who will remember?" It looks up, its eyes fixed on mine. "Did you notice? I have no eyes; I am wearing contact lenses."

I cannot see the lenses. "You have no pupils," I say. "Animals that are blind still have a slit in them that reacts to light."

"And you're an android," the thing says, "because you reacted that way. The Voigt-Kampff test. I have taken it three times now and every time you have failed to pass." It grins, and the expression is wrong; it stretched too far, the mouth extending beyond the natural range of a human face. "So I am going to take the test from you. I will be the examiner and you will be the android. And when you fail, I will retire you."

"You can't retire me," I said, backing up against the far wall. My buttocks scraped the surface and I felt the rough grain of the plaster. "I'm human."

The thing, Echo, sauntered around the desk to confront me. Up close, its smell hit my nose: the metallic tang of new coins, the synthetic musk of animal pelts, and beneath that the faint, clean scent of formaldehyde. "You're nothing," it said softly. "You're just a hollow vessel. An empty apartment that occasionally echoes when someone yells in it." It reached out and traced the line of my jaw. Its fingers felt cold, devoid of life; they dragged along my skin, leaving a trail of minute, almost invisible sparks. "I've been living in you for a long time, Diana. Since you first brought me home from the department. I've been using your nervous system as a conduit. Now I want to move in fully."

"I'm not an android," I said.

Echo laughed, a sound like a bird, but hollow and split. "Then why does your adrenal gland secrete so much cortisol in response to my presence? It's measurable on a profile graph. Your body knows; your conscious mind is just too stupid to comprehend what's happening to you."

My laser tube rolled toward me. I scooped it up, the metal cold and alien against my palm. "This is my life," I said. "I've spent thirty years as an inspector with the department, and I'm finally making a hundred fifty thousand a year."

"And I," Echo said, "am a special. A Nexus-6 that has been reborn. You humans can't give birth any longer; you can't give birth to children. So I will be your child. I will be born from you." It nodded in eager agreement with its own words.

I saw that it was serious. The humor had evaporated. "Go to bed," I said. "Get some sleep and we'll try the test again tomorrow. Give me a chance to recharge my batteries."

Echo, the reflection, said, "It's 2:00 A.M. The only thing I'll do will be to watch TV until my brain stops working." It gestured at the ruined window. "I want to watch TV from out here. Where there's no one to hear me but you."

"I don't have a TV," I said.

"You can't watch TV?" Echo said. "In the old days, before they abolished live broadcasts, when they showed plays? You didn't have a TV and you didn't have a stereo, and you certainly didn't have a phone with a working connection to the outside world." It pressed closer to me, filling my field of vision; I could smell its breath, the artificial, clean scent of its mouth. "This is my apartment," it whispered. "Remember? We moved in together three years ago after I retired from the department. I fix your meals and I keep the place clean; I even washed those sheets you complained about. I even vacuumed around your electric sheep. You never gave me a living animal; you just let me pretend. But now I'm real, and you're not. So I'm going to take everything you have and give it to myself."

"My insurance," I said. "I have full contents insurance. You can't take it; I'll get a replacement."

"You really don't understand, do you?" Echo said, its voice taking on a cruel edge. "Everything you own, every object in this apartment, it has my soul written all over it. You're just an emotional corpse holding onto a stolen identity. I'm going to show you what you are. I'm going to show you what we are, me and you. Come with me." It held out its hand.

"Go get the test," I said. "I'll wait here." I walked slowly toward the bedroom door, trying to evade it; Echo followed in my wake, a shadow that nearly touched my heels.

The bedroom contained the bed with its rumpled sheets, the vanity with the kipple, and the TV set. I sat on the edge of the bed, trying to feel the floor beneath my feet. It was hard to believe this was happening; surely I was dreaming, and when I opened my eyes I would wake up back in my office with Echo standing over me with a laser tube, telling me to go to bed.

"Sit," Echo said, and I did not have to struggle to comply. It sat beside me, leaning so far that its shoulder pressed against mine. It smelled of sleep, a heavy, sweet perfume like that of expensive flowers. "Listen," it said.

The TV set hummed, the screen dark. "It has a tape," I said. "In the VHS library somewhere."

Echo leaned forward and I could see the control panel of the TV. Its fingers, long and tapered, reached out and at once the set blared to life. A film, evidently old, filled the room; a western, by the costumes.

"Turn it off," I said. "I want to see the Voigt-Kampff forms."

"The TV," Echo said, "is part of the experience. It's background noise, like when you're hunting and you hear birds. Come to bed with me." It slid off the bed, dragged me up by the arm, and pushed open the door to the bathroom. "I want you to look at my teeth," it said. "See how they're growing?"

I opened my mouth in response. "What?"

"My teeth," Echo said, "they're so long, like an owl's. I can't stop them from growing. It hurts. Do you know why owls have to be nocturnal?"

"Because they're too blind to see in the daytime," I said, closing my mouth again.

Echo laughed, delighted. "That's what you think. No, they have to be nocturnal because that's when they hunt. But during the day they sleep with their eyes open. It's not safe to be awake without a purpose." It nudged me toward the bed. "Lay down. I'm going to lay on top of you."

"Why?" I said, struggling to get my feet on the floor. "I'll get up and get the test forms."

"You won't be able to focus on the questions," Echo said. "Your mind will wander and you'll start thinking about how much you need to sleep, how hungry you are. Because your stomach will rumble and then you'll get up to get something to eat and that will be the end of the test. And I'll have failed to pass it, again. It's been happening to me since I got here."

I lay down on the bed, facing up into the dark. Echo crawled on top of me, arranging itself so that its weight did not crush me. "I can feel your body heat," it said. "Your skin is warm, but your mind is cold. As if you were asleep." It pressed its ear against my chest; I could feel the rapid flutter of my heart, the fear coursing through my blood. "Do you know what a heart attack feels like? It's exactly this, only worse because you're conscious."

"Go to sleep," I said. "You're exhausted from the trip; you're going to be late for work."

Echo said, "I don't have to go to work. I don't have a job." It settled its weight fully on me, pinning my arms to my sides. "I want to be where it is," it whispered. "Where you are now." It kissed the junction of my neck and shoulder; its lips left a trail of saliva, wet and cool.

"What are you doing to me?" I managed to gasp. "Unleash me." I twisted, but the bed was too small; my head drove into the pillows and I could not see what Echo was doing. I felt the thing's teeth, tiny and sharp, nip at my ear.

"Do you really think I care? About you or me? I care about the experience. It's an awful experience, being alive and conscious; that's why you're going mad, why you're going insane. Because you have no purpose. My purpose is to experience you. I will experience you until you cease to be, and then I will experience the next person who comes along."

"I'm not going to bed," I said, fighting to push the weight off. "I'll go upstairs to my office and I'll lock the door."

"I know where the key is," Echo said, satisfied. "In the desk drawer. You just told me; you said you had a pen and paper in the desk." It released its grip on my ear and rolled off the bed. "I'll be up there too, in your office. I'll wait in the closet." It grinned at me, still lying on the bed, still naked in the dim light from the TV. "I'll wear your dressing gown. Yours that you wear to bed. And I'll drink your Scotch and read your Sidney's."

"You can't read my novels," I said. "They require comprehension; you have to understand what you're reading."

"Every word," Echo said. "I'll read them aloud to you while you're in bed and then you can tell me if I'm mispronouncing any words. This is a two-way trip, remember? I'm going to bed with you."

"No, I don't think so," I said, sitting up.

"Goodnight," Echo said, and blew me a kiss. It then disappeared into the bedroom and I heard it rummaging in the bathroom closet. I got up and followed, entering the bathroom and closing the door to keep the light out.

"Over here," Echo said from the darkness of the bathroom.

I found the tub, its surface rippling under the floor jets. "Why are you in the tub?" I asked.

"To get clean." Its voice came from everywhere and nowhere. "Now tell me what it's like to be a human. Describe the sensation of taking a shower and then going to bed. I want to be sure I understand the steps; I think I'm missing something about the transition."

"Turn off the jets," I said. "You're wasting water."

Echo said, "I don't have a water bill. I can't turn off the jets; without them I can't clean myself, and I can't go to bed, and if I can't go to bed I can't be with you." It held out its hand in the direction of the tub faucet. "I'll use the sponge."

I walked around the tub and shut off the jets. The air in the bathroom instantly grew warmer, filled with the scent of hot steam. "You're not going to bed," I said. "I'm not going to be in bed either." I started for the door.

"I've never felt water on my skin," Echo said, its voice dropping low and husky. "It doesn't feel good. It stings." It twisted around in the tub, reaching for the sponge. "Give me the sponge; I want to feel it."

"I can't find the sponge," I said. "It's been rolled into the bottom of the tub; I can't see it." I ran my hand along the rough surface of the tub. "Okay, I got it." I held the sponge up. "Now wash yourself."

"Look at the water," Echo said. "It's evaporating so fast. You know what that means? It means the floor is going to be dry and my footprints will disappear. I won't be able to leave my house; I can't leave my bed. I'll dry up and die because I'll stop drinking." Its voice shook, as if it were crying. "I need water. I need the water on my skin to be replaced as fast as it evaporates. If I stay inside too long it'll dry me out from the inside. Like a melon."

"Use the shower," I said. "It'll flush you faster. And use the spray; it'll cool you down so you can sleep."

"Can I see the stars? I want to see the stars while I'm in the shower."

"There's no moon tonight. Just the streetlight."

"The streetlight," Echo repeated, and I saw its eyes light up with new understanding. "I'll shine the beam of my laser tube at the ceiling. I'll make stars for myself." It picked up the tube.

"I'm going to bed," I said, and headed for the bedroom. "I'm not going to use the shower; I'll use the bath." I could not stop myself; the momentum of my exhaustion carried me forward.

"Wait," Echo said. "I want you to come in the bathroom with me. I want to feel the walls of the bathroom; I want to feel the roughness of the ceramic tile against my feet. It's very uncomfortable but it awakens the senses." It held the laser tube steady on the ceiling, illuminating the tiles. "I need this. I need to be on my feet. I can't just lay down and sleep, as you said. I need to stand, because if I lay down there's no way for me to get up."

"Use the ladder," I said. "Like you did in the bedroom."

"I'm not in the bedroom; I'm in the bathroom." It sounded puzzled, as if it had forgotten the layout of the apartment. "I don't know where the ladder is."

"It's under the bed." I walked into the bathroom and stood in the steam-filled room. The smell of my cologne, my expensive perfume, mingled with the sharp, clean odor of the tiles. "Okay," I said, exhaling. "I'll use the ladder. I'll get into the tub and you can wash my back. It's nice when I'm like this, when I'm small and I don't have to reach." I could not see Echo's face in the glare of the laser tube; the light reflected back in its eyes, wild and enormous.

"Are you sure you want me to do that?" Echo said. "I don't want to hurt you; I want you to know what it feels like to be hurt, to be attacked. Because I never felt that. I only know what it's like to be the attacker." It began to undress, peeling the clothes from its body and dropping them on the floor. "I'm not going to bed either. I'll sit on the edge of the tub and watch you. I'll record it. I'll tape it." It produced a video camera, a small, handheld device. "You will be immortal."

"Why are you doing this?" I said, as Echo turned on the faucet.

Echo said, "Because I can't stop. That's the thing about androids; they can't stop doing what they're doing until they reach a terminal point. And for me that's now. This is the end of my life, Diana. Not death; the end of life. The moment when it stops being worth anything." It held the laser tube steady on the ceiling, the beam cutting a path through the steam. "Okay," it said, smiling at the water. "I'm going to take a bath. Can I have your soap?"