Instructions to use lordx64/Qwen3.6-35B-A3B-Kimi-K2.6-Reasoning-Distilled with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use lordx64/Qwen3.6-35B-A3B-Kimi-K2.6-Reasoning-Distilled with Transformers:

# Use a pipeline as a high-level helper
from transformers import pipeline

pipe = pipeline("text-generation", model="lordx64/Qwen3.6-35B-A3B-Kimi-K2.6-Reasoning-Distilled")
messages = [
    {
        "role": "user",
        "content": [
            {"type": "image", "url": "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/p-blog/candy.JPG"},
            {"type": "text", "text": "What animal is on the candy?"}
        ]
    },
]
pipe(text=messages)

# Load model directly
from transformers import AutoProcessor, AutoModelForMultimodalLM

processor = AutoProcessor.from_pretrained("lordx64/Qwen3.6-35B-A3B-Kimi-K2.6-Reasoning-Distilled")
model = AutoModelForMultimodalLM.from_pretrained("lordx64/Qwen3.6-35B-A3B-Kimi-K2.6-Reasoning-Distilled")
messages = [
    {
        "role": "user",
        "content": [
            {"type": "image", "url": "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/p-blog/candy.JPG"},
            {"type": "text", "text": "What animal is on the candy?"}
        ]
    },
]
inputs = processor.apply_chat_template(
	messages,
	add_generation_prompt=True,
	tokenize=True,
	return_dict=True,
	return_tensors="pt",
).to(model.device)

outputs = model.generate(**inputs, max_new_tokens=40)
print(processor.decode(outputs[0][inputs["input_ids"].shape[-1]:]))

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use lordx64/Qwen3.6-35B-A3B-Kimi-K2.6-Reasoning-Distilled with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "lordx64/Qwen3.6-35B-A3B-Kimi-K2.6-Reasoning-Distilled"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "lordx64/Qwen3.6-35B-A3B-Kimi-K2.6-Reasoning-Distilled",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/lordx64/Qwen3.6-35B-A3B-Kimi-K2.6-Reasoning-Distilled

SGLang

How to use lordx64/Qwen3.6-35B-A3B-Kimi-K2.6-Reasoning-Distilled with SGLang:

Install from pip and serve model

# Install SGLang from pip:
pip install sglang
# Start the SGLang server:
python3 -m sglang.launch_server \
    --model-path "lordx64/Qwen3.6-35B-A3B-Kimi-K2.6-Reasoning-Distilled" \
    --host 0.0.0.0 \
    --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "lordx64/Qwen3.6-35B-A3B-Kimi-K2.6-Reasoning-Distilled",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker images

docker run --gpus all \
    --shm-size 32g \
    -p 30000:30000 \
    -v ~/.cache/huggingface:/root/.cache/huggingface \
    --env "HF_TOKEN=<secret>" \
    --ipc=host \
    lmsysorg/sglang:latest \
    python3 -m sglang.launch_server \
        --model-path "lordx64/Qwen3.6-35B-A3B-Kimi-K2.6-Reasoning-Distilled" \
        --host 0.0.0.0 \
        --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "lordx64/Qwen3.6-35B-A3B-Kimi-K2.6-Reasoning-Distilled",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Unsloth Studio

How to use lordx64/Qwen3.6-35B-A3B-Kimi-K2.6-Reasoning-Distilled with Unsloth Studio:

Install Unsloth Studio (macOS, Linux, WSL)

curl -fsSL https://unsloth.ai/install.sh | sh
# Run unsloth studio
unsloth studio -H 0.0.0.0 -p 8888
# Then open http://localhost:8888 in your browser
# Search for lordx64/Qwen3.6-35B-A3B-Kimi-K2.6-Reasoning-Distilled to start chatting

Install Unsloth Studio (Windows)

irm https://unsloth.ai/install.ps1 | iex
# Run unsloth studio
unsloth studio -H 0.0.0.0 -p 8888
# Then open http://localhost:8888 in your browser
# Search for lordx64/Qwen3.6-35B-A3B-Kimi-K2.6-Reasoning-Distilled to start chatting

Using HuggingFace Spaces for Unsloth

# No setup required
# Open https://huggingface.co/spaces/unsloth/studio in your browser
# Search for lordx64/Qwen3.6-35B-A3B-Kimi-K2.6-Reasoning-Distilled to start chatting

Load model with FastModel

pip install unsloth
from unsloth import FastModel
model, tokenizer = FastModel.from_pretrained(
    model_name="lordx64/Qwen3.6-35B-A3B-Kimi-K2.6-Reasoning-Distilled",
    max_seq_length=2048,
)

Docker Model Runner
How to use lordx64/Qwen3.6-35B-A3B-Kimi-K2.6-Reasoning-Distilled with Docker Model Runner:
```
docker model run hf.co/lordx64/Qwen3.6-35B-A3B-Kimi-K2.6-Reasoning-Distilled
```

lordx64 commited on Apr 27

Commit

0bf86f7

verified ·

1 Parent(s): 6c94ebb

Eval: add base MATH-500=53% + TL;DR explaining what distillation actually does (style transfer, not capability gain)

Browse files

Files changed (1) hide show

README.md +17 -3

README.md CHANGED Viewed

@@ -145,12 +145,21 @@ The initial plan was full LoRA including the MoE expert FFNs (`gate_proj/up_proj
 Methodology: vLLM + lm-eval-harness with a custom `<think>`-stripping wrapper, `max_gen_toks=16384` to allow full Kimi-style reasoning chains before answer extraction. Each model evaluated under identical conditions on a single H200. See [`training/eval.py`](https://github.com/lordx64/distillation/blob/main/training/eval.py).
 ### Head-to-head: Kimi-Distill vs Base
 | Benchmark | Setup | Base Qwen3.6-35B-A3B | **Kimi-Distill (this model)** | Δ |
 |---|---|---:|---:|---:|
 | **GSM8K** | 8-shot CoT, 300 examples, strict-match | 64.00% | **92.67%** | **+28.67 pp** ✅ |
-| **MATH-500** | 0-shot, 100 problems, math_verify | _running_ | **47.00%** | _pending base re-run_ |
 | **GPQA Diamond** | 0-shot CoT, 198 problems, flex-extract | 79.29% | 75.25% | -4.04 pp |
 | **MMLU-Pro math** | 5-shot, custom-extract | 27.20% | **64.80%** | **+37.60 pp** ✅ |
 | **MMLU-Pro CS** | 5-shot, custom-extract | 20.49% | **61.46%** | **+40.97 pp** ✅ |
@@ -159,9 +168,14 @@ Methodology: vLLM + lm-eval-harness with a custom `<think>`-stripping wrapper, `
 | MMLU-Pro overall | 5-shot, custom-extract | 6.35% | 14.67% | +8.32 pp (extractor-affected for both) |
 | AIME 2024 / 2025 | 0-shot, 30 problems, strict-match | 0.00% | 0.00% | extractor format issue (see note) |
-The headline: **on every benchmark where the extractor produces clean numbers, the Kimi-distill clearly outperforms the base** — most dramatically on GSM8K (+28.67pp), MMLU-Pro Math (+37.60pp), and MMLU-Pro Computer Science (+40.97pp). The distillation transferred Kimi K2.6's verbose reasoning style robustly enough that the student emits `<think>` blocks unconditionally, even on fewshot prompts that don't model the pattern, while the base imitates the fewshot format and skips reasoning.
-GPQA Diamond is the one benchmark where the base edges out the distill (-4.04 pp). This is consistent with distillation transferring reasoning *style* but not adding factual knowledge — GPQA is largely a knowledge benchmark and the base's STEM coverage is what answers most questions.
 ### Notes on the methodology issues

 Methodology: vLLM + lm-eval-harness with a custom `<think>`-stripping wrapper, `max_gen_toks=16384` to allow full Kimi-style reasoning chains before answer extraction. Each model evaluated under identical conditions on a single H200. See [`training/eval.py`](https://github.com/lordx64/distillation/blob/main/training/eval.py).
+> **TL;DR — what this distillation actually does, based on the head-to-head data below:**
+>
+> **The distillation makes the model unconditionally think, but doesn't improve raw reasoning capability over the base.**
+>
+> - On benchmarks where the **base fails to invoke its own thinking** (GSM8K under fewshot pattern, most MMLU-Pro subjects): distill **+28 to +40 pp**.
+> - On benchmarks where the **base already thinks natively** (MATH-500, GPQA Diamond): distill **−4 to −6 pp** (a small style-imitation cost).
+>
+> Use this model when you want predictable `<think>`-block reasoning on prompts that don't trigger the base's thinking mode (fewshot evals, plain Q&A, latency-sensitive deploys with fixed prompt templates). Use the base directly when you can already prompt it for zero-shot CoT.
 ### Head-to-head: Kimi-Distill vs Base
 | Benchmark | Setup | Base Qwen3.6-35B-A3B | **Kimi-Distill (this model)** | Δ |
 |---|---|---:|---:|---:|
 | **GSM8K** | 8-shot CoT, 300 examples, strict-match | 64.00% | **92.67%** | **+28.67 pp** ✅ |
+| **MATH-500** | 0-shot, 100 problems, math_verify | 53.00% | 47.00% | -6.00 pp |
 | **GPQA Diamond** | 0-shot CoT, 198 problems, flex-extract | 79.29% | 75.25% | -4.04 pp |
 | **MMLU-Pro math** | 5-shot, custom-extract | 27.20% | **64.80%** | **+37.60 pp** ✅ |
 | **MMLU-Pro CS** | 5-shot, custom-extract | 20.49% | **61.46%** | **+40.97 pp** ✅ |
 | MMLU-Pro overall | 5-shot, custom-extract | 6.35% | 14.67% | +8.32 pp (extractor-affected for both) |
 | AIME 2024 / 2025 | 0-shot, 30 problems, strict-match | 0.00% | 0.00% | extractor format issue (see note) |
+### What the data is actually telling us
+Two patterns, depending on whether the base model invokes its own thinking under the eval prompt:
+- **Where the base fails to invoke thinking** (GSM8K under fewshot pattern, most MMLU-Pro subjects), the distillation wins by a wide margin: **GSM8K +28.67 pp**, MMLU-Pro Math **+37.60 pp**, MMLU-Pro CS **+40.97 pp**. The Kimi-distill emits `<think>` blocks unconditionally — it doesn't follow the fewshot pattern; it reasons regardless. That's a real prompt-robustness gain.
+- **Where the base already thinks natively** (MATH-500, GPQA Diamond), the distill is **slightly worse**: -6.00 pp on MATH-500, -4.04 pp on GPQA. Style imitation incurs a small cost when the base was already doing the right thing.
+So the honest framing: this distillation **transfers Kimi K2.6's verbose reasoning style and makes the model think unconditionally**. It does not add raw reasoning capability or factual knowledge over the base. Use the Kimi-distill when you want predictable thinking on prompts that don't trigger the base's own thinking mode (fewshot evals, plain Q&A patterns); use the base directly when you can already prompt it correctly (zero-shot CoT on math/STEM).
 ### Notes on the methodology issues