Alibaba releases Qwen3.6-35B-A3B, an open multimodal MoE model built for agentic coding

Alibaba's Qwen team has released the weights for Qwen3.6-35B-A3B, a new multimodal model built on a sparse MoE architecture. It has 35 billion total parameters, but only 3 billion activate during inference. The focus is on agentic coding, tool use and multimodal understanding.

What was released

Qwen3.6-35B-A3B is the first open-weight model in the Qwen3.6 family, following the launch of Qwen3.6-Plus. It ships under the Apache 2.0 license. You can self-host it from Hugging Face and ModelScope or access it through the Alibaba Cloud Model Studio API. It isn't a text-only model: a built-in vision encoder lets it accept images and video. The native context window is 262,144 tokens and can be extended to roughly 1.01 million.

The core idea behind the release is strong performance from a small number of active parameters. The model holds 35 billion parameters, but only about 3 billion do work at each step. According to the model card, the architecture uses 256 experts, with 8 routed experts and 1 shared expert active at a time. In practice, that makes inference cheaper than with large dense models. Qwen3.6 runs in thinking mode by default, but it can also answer directly without intermediate reasoning.

A bet on code

Qwen explicitly pitches this release as a model for agentic coding, not just another general-purpose chatbot. The developers say Qwen3.6-35B-A3B handles frontend tasks, repository navigation and multi-step tool use better than before. The model integrates with Qwen-Agent, OpenClaw, Qwen Code and even Claude Code via compatible APIs. For long sessions, a dedicated preserve_thinking feature keeps reasoning chains from earlier messages, so the agent doesn't have to rebuild context from scratch at every step.

• Tool calling and agentic pipelines
• Multi-file repository analysis
• Frontend code generation and editing
• Long iterative sessions that retain reasoning context

According to Qwen, the model looks strongest on coding and agentic tests. It scores 73.4 on SWE-bench Verified, 51.5 on Terminal-Bench 2.0, 29.4 on NL2Repo and 1397 on the internal QwenWebBench. That's well ahead of Qwen3.5-35B-A3B and, on several tasks, better than the larger dense Qwen3.5-27B.

In other words, Qwen is trying to show that an open-weight MoE model can be useful for more than local chat: it can also handle full dev workflows that require tools, memory of previous steps and work across an entire codebase.

Multimodal without compromises

Qwen also puts particular emphasis on vision and multimodal reasoning. According to Qwen's tables, the model scores 85.3 on RealWorldQA, 92.8 on MMBench EN, 89.9 on OmniDocBench1.5 and 81.9 on CC-OCR. The spatial understanding results are even more interesting: 92.0 on RefCOCO and 50.8 on ODInW13.

The video numbers are strong too: 83.7 on VideoMMMU and 86.2 on MLVU. For a model with 3 billion active parameters, that's a serious claim to versatility rather than a narrow focus on code alone.

On the practical side, Qwen3.6-35B-A3B drops into familiar inference stacks such as vLLM and SGLang, with modes for tool use and a language-only launch if you need to free up memory. Qwen's examples run the full 262K context on eight GPUs, but the team separately advises not going below 128K where possible if thinking capabilities matter. For teams that want to host the model themselves and avoid depending on closed SaaS, this already looks less like an experiment and more like a working option.

What it means

Qwen keeps pushing the open-weight market toward more practical models: not maximum size for its own sake, but a balance of inference cost, long context, multimodality and real value in development work. If the claimed results hold up in real-world use, Qwen3.6-35B-A3B will be one of the most interesting open options for teams that need an AI assistant for code, documents, images and agentic tasks without being locked into closed platforms.