--- library_name: transformers license: apache-2.0 pipeline_tag: text-generation language: - en - zh tags: - qwen3_5_moe - qwen-agentworld - world-model - agentic-environment-simulation - mtp - apex - quantization - gguf - multimodal base_model: - Qwen/Qwen-AgentWorld-35B-A3B ---
APEX-I MTP GGUF Apache-2.0

Qwen-AgentWorld-35B-A3B-MTP-APEX

English | 中文 README

AgentWorld world model · native MTP injected · APEX-I GGUFs

Important Warning

Qwen-AgentWorld is designed as an agentic environment simulation / language world model. It is intended to simulate agent-environment interactions and predict environment states from actions and history. It is not recommended as a general daily assistant, coding model, or autonomous Agent backbone unless you specifically need this world-model behavior.

About This Release

This package contains GGUF releases of Qwen-AgentWorld-35B-A3B, a Qwen3.5 MoE world model for agentic environment simulation.

The model is the original Qwen release with no LoRA or abliteration applied. One native MTP layer from a compatible Qwen3.5-35B-A3B MTP shard was injected into the safetensors checkpoint, the result was converted to BF16 GGUF, and then quantized into APEX-I tiers using a model-specific imatrix.

Model Details
Base modelQwen/Qwen-AgentWorld-35B-A3B
ArchitectureQwen3.5 MoE
Parameters35B total, around 3B active per token
Experts256 routed experts, 8 active per token
Context262,144 tokens
MTP1 injected MTP layer, 785 tensors
LicenseApache-2.0
MTP Extraction And Injection

The MTP layer was extracted from a compatible Qwen3.5-35B-A3B MTP shard and injected into the AgentWorld safetensors model.

Tensor prefixmtp.*
Tensor count785
Injected shardmodel-00022-of-00022.safetensors
BenchLocal Results (I-Compact, thinking)
ModelTCBFHACeilingPracticalRetry
Qwen-AgentWorld-35B-A3B, thinking100878188.587.51
Qwen3.6-35B thinking baseline97888589.581.58

Thinking mode is strongly recommended for this model family.

Usage

llama.cpp server

hf download SC117/Qwen-AgentWorld-35B-A3B-MTP-APEX-GGUF --include "*.gguf" --local-dir ./models/qwen-agentworld-mtp-apex ./llama-server -m ./models/qwen-agentworld-mtp-apex/Qwen-AgentWorld-35B-A3B-MTP-APEX-I-Compact.gguf -ngl 99 -c 131072

vision + text

./llama-server -m ./models/qwen-agentworld-mtp-apex/Qwen-AgentWorld-35B-A3B-MTP-APEX-I-Compact.gguf --mmproj ./models/qwen-agentworld-mtp-apex/mmproj-f16.gguf -ngl 99 -c 131072

APEX Quantization Tiers
FileSizeProfileBest For
*-APEX-I-Balanced.gguf24.33 GBI-BalancedHighest-retention all-rounder
*-APEX-I-Quality.gguf21.87 GBI-QualityQuality-focused, IQ4_XS middle experts
*-APEX-I-Compact.gguf16.14 GBI-CompactRecommended quality/size ratio
*-APEX-I-Mini.gguf13.29 GBI-MiniSmallest practical tier

Quantized with a model-specific imatrix generated from cleaned AgentWorld calibration text.

## Links - **Original Model**: https://huggingface.co/Qwen/Qwen-AgentWorld-35B-A3B - **Qwen-AgentWorld Project**: https://github.com/QwenLM/Qwen-AgentWorld - **APEX Quantization**: https://github.com/mudler/apex-quant - **BenchLocal Thinking**: https://scorp1o117.github.io/benchlocal-results/models/qwen-agentworld-35b-thinking.html ## Citation ```bibtex @misc{qwen-agentworld-2026, title = {Qwen-AgentWorld: Language World Models for General Agents}, author = {Qwen Team}, year = {2026}, url = {https://github.com/QwenLM/Qwen-AgentWorld} } ```