Qwen3.6-35B-A3B Release Notes
Qwen has open-sourced Qwen3.6-35B-A3B, a sparse mixture-of-experts (MoE) model designed for high-efficiency agentic coding and multimodal reasoning. The model features 35 billion total parameters but only 3 billion active parameters, allowing it to rival the performance of much larger dense models like Qwen3.5-27B and Gemma4-31B while maintaining a significantly smaller active compute footprint.
Agentic Coding and Reasoning Performance
Qwen3.6-35B-A3B outperforms its predecessor, Qwen3.5-35B-A3B, and competes with larger dense models across several coding and agentic benchmarks. With only 3B active parameters, it achieves superior results in agentic coding tasks compared to the dense 27B-parameter Qwen3.5-27B.
Coding Agent Benchmarks
- SWE-bench Verified: 73.4 (compared to 75.0 for Qwen3.5-27B and 70.0 for Qwen3.5-35B-A3B).
- Terminal-Bench 2.0: 51.5 (surpassing Qwen3.5-27B's 41.6).
- Claw-Eval Avg: 68.7 (surpassing Qwen3.5-27B's 64.3).
- NL2Repo: 29.4 (surpassing Qwen3.5-27B's 27.3).
- QwenWebBench: 1397 (surpassing Qwen3.5-27B's 1068).
General Agent and Knowledge Benchmarks
- MCPMark: 37.0 (compared to 36.3 for Qwen3.5-27B).
- AIME26: 92.7 (compared to 92.6 for Qwen3.5-27B).
- GPQA: 86.0 (compared to 85.5 for Qwen3.5-27B).
Multimodal Capabilities
Qwen3.6-35B-A3B is natively multimodal, delivering perception and reasoning capabilities that match or exceed Claude Sonnet 4.5 on various vision-language benchmarks. Its performance is particularly strong in spatial intelligence.
Vision-Language Benchmarks
- Spatial Intelligence: Achieves 92.0 on RefCOCO and 50.8 on ODInW13.
- STEM and Puzzle: 81.7 on MMMU and 86.4 on Mathvista(mini).
- General VQA: 85.3 on RealWorldQA and 92.8 on MMBench EN-DEV-v1.1.
- Document Understanding: 89.9 on OmniDocBench1.5 and 81.9 on CC-OCR.
- Video Understanding: 83.7 on VideoMMMU and 86.2 on MLVU.
Integration and Deployment
Qwen3.6-35B-A3B is available as open weights on Hugging Face and ModelScope, and via the Alibaba Cloud Model Studio API (as qwen3.6-flash).
API Features and Compatibility
- Thinking Preservation: The API supports the
preserve_thinkingfeature, which keeps thinking content from preceding turns in messages, a feature recommended for agentic tasks. - Protocol Support: Alibaba Cloud Model Studio supports OpenAI-compatible chat completions and responses APIs, as well as an API interface compatible with Anthropic.
Third-Party Coding Assistant Integration
- OpenClaw: Compatible with the self-hosted open-source AI coding agent OpenClaw.
- Qwen Code: Deeply optimized for the Qwen series and available as a terminal-based AI agent.
- Claude Code: Supported via the Anthropic API protocol compatibility of Qwen APIs.
Summary of Model Specifications
| Feature | Specification |
|---|---|
| Total Parameters | 35 Billion |
| Active Parameters | 3 Billion |
| Architecture | Sparse Mixture-of-Experts (MoE) |
| Availability | Open Weights (Hugging Face, ModelScope) / API (qwen3.6-flash) |