Advanced Coding, Reasoning & Multimodal Understanding with Built-In Thinking
Alibaba's latest open-weight multimodal AI. Purpose-built for agentic coding, complex reasoning, and real-world software development, with frontier-level performance at a fraction of the compute cost of comparable models.
What it does
Qwen3.6-35B-A3B is the newest model in Alibaba's Qwen family, built in direct response to developer community feedback following the Qwen3.5 release. While it is a general-purpose multimodal AI capable of understanding text, images, and video, its standout strength is agentic coding, the ability to autonomously plan, write, debug, and execute complex software tasks across entire repositories and frontend workflows. A key architectural advantage is its Mixture-of-Experts (MoE) design: despite having 35 billion total parameters, it only activates 3 billion at a time during inference, meaning it runs with the speed and cost of a much smaller model while retaining the intelligence of a much larger one.
Problem it solves
- Agentic coding at the frontier – Leads open-source models on SWE-bench Verified (73.4%), the most rigorous real-world software engineering benchmark, and tops the QwenWebBench frontend coding leaderboard with a score of 1,397
- Cost-efficient frontier intelligence – The MoE architecture activates only 3B parameters per inference pass, making it dramatically cheaper and faster to run than dense 35B models of equivalent quality
- Complex, multi-step reasoning – Built-in thinking mode allows the model to reason step-by-step before responding, significantly improving accuracy on hard problems in math, science, and coding
- Thinking preservation – A new capability allowing the model to retain and build on its reasoning from earlier in a conversation, reducing redundant thinking and improving consistency across long multi-turn workflows
- Ultra-long context – A native 262K token context window (extensible to over 1 million tokens) allows it to work across entire codebases, long documents, or extended multi-turn conversations in one pass
- Agentic tool use – Native function calling and MCP (Model Context Protocol) support enables it to connect to external tools, APIs, file systems, and services to act autonomously on real-world tasks
Input/Output
- Input:
- Text – Questions, instructions, code, documents, and multi-turn conversations
- Images – Photos, screenshots, charts, diagrams, and documents
- Advanced:
- Max New Tokens: Controls the maximum length of the model's response. For most tasks, 32,768 tokens is recommended. For complex math or coding competition problems, up to 81,920 tokens is advised to give the model sufficient space to reason thoroughly.
- Temperature: Controls how creative or predictable the response is. Set to 0.6 for precise coding tasks where accuracy is critical. Use 1.0 for general tasks and open-ended reasoning where more varied responses are beneficial.
- Nucleus sampling probability threshold: Sets a probability threshold that limits which words the model considers for its next response. For example, a Top-P of 0.95 means the model only picks from the most likely words that together add up to 95% probability — filtering out unlikely or off-topic choices. Lower values make responses more focused and predictable; higher values allow more variety. Recommended: 0.95 for thinking mode, 0.8 for standard instruct mode.
- Top-k sampling: Limits the model to choosing only from the top K most likely next words at each step. For example, Top-K of 20 means the model only considers the 20 most probable options, ignoring everything else. Works alongside Top-P to keep responses coherent and on-topic. Recommended: 20 across all modes.
- Enable thinking mode for chain-of-thought reasoning before answering: When turned on, the model reasons through the problem step-by-step before delivering its final answer — similar to showing its work. This significantly improves accuracy on complex tasks like coding, mathematics, and multi-step reasoning. When turned off, the model responds directly and more quickly, which is better suited for simple, conversational, or time-sensitive tasks. On by default in Qwen3.6.
- Output:
- Generated text, analysis, code, summaries, or structured data
- Supports a built-in thinking mode where the model reasons step-by-step before delivering its final answer, particularly valuable for complex coding, math, and reasoning tasks
Accuracy & Speed
Benchmark Performance
Coding & Agentic Tasks
Benchmark | Qwen3.6-35B-A3B | What it measures |
SWE-bench Verified | 73.4% | Real-world software engineering |
SWE-bench Multilingual | 67.2% | Cross-language software engineering |
Terminal-Bench 2.0 | 51.5% | Autonomous terminal / command-line tasks |
QwenWebBench | 1,397 (ELO) | Frontend code generation |
NL2Repo | 29.4% | Natural language to full repository generation |
Reasoning & Knowledge
Benchmark | Qwen3.6-35B-A3B | What it measures |
AIME 2026 | 92.7% | Advanced mathematics |
GPQA Diamond | 86.0% | Expert-level science reasoning |
MMLU-Pro | 85.2% | Broad expert knowledge |
LiveCodeBench v6 | 80.4% | Real-world coding ability |
Vision & Multimodal
Benchmark | Qwen3.6-35B-A3B | What it measures |
MMMU | 81.7% | Multimodal academic understanding |
OmniDocBench 1.5 | 89.9% | Document reading and extraction |
VideoMMMU | 83.7% | Video understanding |
RealWorldQA | 85.3% | Real-world visual question answering |
Technical Details
Architecture | Mixture-of-Experts (MoE) — 35B total parameters, only 3B activated per inference |
Context window | 262,144 tokens natively; extensible to 1,010,000 tokens |
Modalities supported | Text, Image, Video |
Thinking mode | Built-in. Reasons step-by-step before responding; can be enabled or disabled per request |
Thinking preservation | New in Qwen3.6: retains reasoning context across multi-turn conversations for more consistent, efficient agentic workflows |
Function calling | Native. Connects to external tools, APIs, and MCP-compatible services |
Supported frameworks | Hugging Face Transformers, vLLM, SGLang, KTransformers |
Compliance & Provenance
Provider | Alibaba (open weights) |
Provider type | GPAI — General-Purpose AI (GPAI) |
License | Apache 2.0 |
EU AI Act risk class | Limited Risk |
Art. 50 transparency | Required — outputs are marked. See AI Policy §2. |
Region availability | Available globally |
Training data summary |
For more on how we classify models and mark outputs, see our AI Policy.
Model Source
- Hugging Face: https://huggingface.co/Qwen/Qwen3.6-35B-A3B
- License: apache-2.0