Alibaba’s Qwen team has released Qwen3.6-35B-A3B, the first open-weight model in the Qwen3.6 family.
According to Qwen’s official release materials, this is a 35 billion parameter model with 3 billion active parameters per token, using a sparse mixture-of-experts design intended to deliver stronger coding performance without the full inference cost of a dense model.
Qwen is positioning the release around agentic coding, not just chat. In its official documentation, the company highlights improvements in frontend workflows, repository-level reasoning, and a feature it calls thinking preservation, designed to retain reasoning context across historical messages during iterative development.
The technical profile is also notable. The official model card describes Qwen3.6-35B-A3B as a causal language model with a vision encoder, with 35B total parameters and 3B activated parameters, a 262,144 token native context window, and support for extension up to 1,010,000 tokens. Release artifacts are published for common inference stacks including Transformers, vLLM, SGLang and KTransformers.
Qwen’s published benchmark table suggests the model is aimed squarely at practical software workflows. On the model card, Qwen3.6-35B-A3B is listed at 73.4 on SWE-bench Verified, 51.5 on Terminal-Bench 2.0, 68.7 on Claw-Eval Avg, and 29.4 on NL2Repo, with Qwen presenting it as an advance over earlier Qwen3.5 open models in coding-agent tasks.
The release also fills an important gap in the broader Qwen lineup. Until now, the headline Qwen3.6 product in Alibaba Cloud Model Studio had been Qwen3.6-Plus, a proprietary API model launched earlier this month and positioned as a native multimodal, 1 million context model for agentic coding. Qwen3.6-35B-A3B gives developers a more open alternative in the same generation, especially for teams that want self-hosting, deeper stack control, or lower-cost experimentation.
The commercial angle is straightforward. Qwen3.6-35B-A3B is not the biggest model in the market, but it may be one of the more interesting recent releases for developers who care about the balance between coding ability, open weights, and efficient inference. In a market crowded with large proprietary models, a 35B model that activates only 3B parameters per token is a practical story, not just a benchmark story.
For teams evaluating Qwen right now, the split is becoming clearer. Qwen3.6-Plus is the premium API-first option. Qwen3.6-35B-A3B is the open-weight option for builders who want to run closer to the metal. That gives the Qwen3.6 family broader reach across both API users and self-hosted developers.