Qwen3-235B (reasoning) removed or otherwise unavailable

jimbo

It seems like the qwen3 reasoning variant was removed from the assistant. I no longer see it listed on the documentation. I still see it in the UI and in archived version of the documentation. If I select it on the assistant, it does not work, but it is still visible even after wiping my browser cache:

I'd really like to see it working again.

Why I think qwen3 reasoning is needed:

With my custom instructions, qwen3 results were more often better than with the ultimate tier research assistant. Primarily because the reasoning reduces the hallucinations very effectively. If I'm not satisfied with the answer, the reasoning lets me see what caused the result to suffer so I can correct my question.

Getting rid of most hallucinations and being able to control the assistant behaviour are in my opinion perhaps the most valuable aspects of an ai model from the user perspective.

Qwen3 reasoning was also labeled as affordable AND private. Privacy is perhaps one of the strongest reasons to become Kagi user. Often it cannot be combined with affordability, but in this case it can be, so I'd find it weird to consider it being missing as intended behaviour.

And sure, I'm using the assistant wrong, I was using a single model for every task, even though I know I should choose the model according to my task needs. But the thing is that I feel like I was very satisfied with the qwen3 reasoning results.

Personally I do not see value for the model speed, since an "almost correct" is rarely useful, ai being confidently wrong wastes my time. I'd really appreciate some tips of an good alternate if "qwen3 reasoning" assistant is really doomed and it not being available isn't a bug. If it is a bug, there's demand for keeping qwen reasoning variants in the future too.

laiz

Hey!
Qwen3-235B-A22B-Thinking-2507 was somewhat unexpectedly discontinued by the upstream provider. It will be replaced in the next update. Our apologies for the sudden disappearance of the model.

jimbo

laiz Thanks for letting me know, I think I'll figure something out in the meantime

jimbo

As for alternatives, I did not realize the qwen3.6 plus is actually a thinking model even though not labeled as one. (EDIT: I was wrong about this, see: https://kagifeedback.org/d/10383-qwen36-plus-outputs-to-thinking) That's perhaps the closest alternative if someone else is looking for one, but it is labeled as bit more expensive.

Other close contenders are perhaps kimi k2.5 and glm 5. I personally spent most of my time with kimi, but even after hours of trying to encourage it to think in similar way as the qwen, I wasn't able to get satisfying results - but it is quite fast. According to the research assistant glm 5.1 and glm 5 are very close in performance but the 5.1 variant is labeled as more expensive, so perhaps not worth it. I did not manage to try it so thoroughly, but it could also be a valid alternative. (EDIT: glm 5.0 reasoning was pretty good cost/quality balance alternative, kimi is meant for different kinds of tasks, so probably not that useful as assistant model)