Alibaba ditches hybrid mode, drops new Qwen3-235B Instruct model

Qwen3-235B-A22B-Instruct-2507 is Alibaba's latest open-weight instruction model, packing 235B total parameters with 22B active parameters.

Scrapping hybrid mode for focused training

Rather than forcing a single model to do both general dialogue and deep reasoning, the Qwen team listened to community feedback and completely ditched its hybrid thinking mode. Moving forward, Instruct and Thinking models are being trained separately to push output quality as high as possible.

According to the team, this iteration is noticeably smarter, broader in knowledge, and much better suited for agent workflows than the prior release. The weights are available on Hugging Face and ModelScope in both full precision and an FP8 version, and you can take it for a spin directly on Qwen Chat. The team noted this is merely an incremental release, with bigger milestones on the horizon.

Why it matters

If you are building self-hosted AI agents or looking for a heavyweight open-source alternative to proprietary APIs, the immediate availability of FP8 weights makes local deployment far more approachable.

Stepping back from the all-in-one hybrid approach shows healthy engineering pragmatism. Splitting instruction and reasoning gives each model a clearer lane, and the real test will be how their upcoming standalone Thinking model measures up against the current reasoning leaders.