Alibaba open-sources Qwen3.6-35B-A3B sparse MoE model

Qwen3.6-35B-A3B is Alibaba's latest open-source sparse Mixture-of-Experts model, built specifically to deliver heavyweight agentic coding on a lightweight compute budget.

Released under the permissive Apache 2.0 license, it represents the first open-weight variant of the Qwen3.6 family following February's Qwen3.5 launch.

What it can do

The model packs 35 billion total parameters, but only routes to 3 billion active parameters per token.

Despite the lean 3B active footprint, the Qwen team claims its agentic coding matches models ten times its active size. It is explicitly tuned for frontend workflows and repository-level multi-file reasoning.

It is also fully multimodal. The model handles visual perception and reasoning tasks, offering both a multimodal thinking mode and a direct non-thinking inference mode.

How to run it

The weights and configs are freely available on Hugging Face and ModelScope under Apache 2.0.

Out of the box, it supports standard serving frameworks like vLLM, SGLang, KTransformers, and Hugging Face Transformers. You can also spin it up immediately using Docker Model Runner.

If you prefer not to host it locally, you can try it via Qwen Studio at chat.qwen.ai or on HuggingChat, with interactive environments ready on Google Colab, Kaggle, and AMD Developer Cloud. Cloud API access via Alibaba's Model Studio is coming soon under the name Qwen3.6-Flash.

Why it matters

If you are an engineer running local coding agents or looking to cut API inference costs, this release is worth an immediate spin. Generating tokens with only 3B parameters active makes inference blazingly fast and dirt-cheap to run at high concurrency.

Alibaba's focus on practical repo-level reasoning gives small teams a real alternative to giant closed models for everyday software maintenance. Just keep in mind that holding 35B total weights still demands enough VRAM to house the model, even if each token only taps a small fraction of it.