Qwen3 235B A22B Thinking 2507: Reasoning Without a Switch

A reasoning model is the wrong default for most requests. It is the right tool for a narrower set: problems where a wrong answer is expensive, the path to the answer has several steps, and you can wait. Qwen3 235B A22B Thinking 2507 is built only for that narrower set. It always reasons, there is no way to turn that off, and its strengths and costs both follow from that.

Qwen released it on 25 July 2025, four days after the non-thinking Qwen3 235B A22B Instruct 2507. Together they replaced the April hybrid model, which switched between modes on request, with two dedicated models. This guide covers the thinking half.

What changed from the hybrid

The architecture is the same as the rest of the 235B line: a mixture-of-experts model with 235B total parameters and 22B active per token, 94 layers, 128 experts with 8 active. The licence is Apache 2.0.

What changed is behaviour. Per the model card, this version is thinking-only, has an increased thinking length compared with the April model, and claims improvements in reasoning, coding, instruction following, tool use and writing. Native context is 262,144 tokens, up from 32,768 for the April original, with a documented configuration for about 1M tokens that the card says needs roughly 1,000 GB of GPU memory.

Some catalogues list a smaller context for this model than the vendor does; OpenRouter, for example, has shown 128,000. Check what your provider actually serves before you design around 256K.

Where the reasoning numbers land

The model card compares it against Gemini 2.5 Pro, o3, o4-mini, Claude Opus 4 in thinking mode and DeepSeek-R1-0528. Its own figures:

• AIME25 — 92.3, just behind o4-mini at 92.7
• HMMT25 — 83.9, ahead of Gemini 2.5 Pro at 82.5
• LiveCodeBench v6 — 74.1, ahead of Gemini 2.5 Pro at 72.5
• GPQA — 81.1, behind Gemini 2.5 Pro at 86.4
• Humanity's Last Exam — 18.2, behind Gemini 2.5 Pro at 21.6
• OJBench — 32.5, behind Gemini 2.5 Pro at 38.9
• TAU2-Telecom — 45.6, well behind o3 at 60.5

The shape is consistent. Competition maths and competitive programming are where it matched or led the closed reasoning models of mid-2025. Broad expert knowledge (GPQA, Humanity's Last Exam) and multi-turn agentic tool use (the TAU benchmarks) are where it trailed the strongest closed models, with o3 ahead of it on every TAU task in the table. These are vendor-run numbers, and the comparison models have all since been superseded, so use them to understand the model's profile rather than its rank today.

Output budget is the real cost

OpenRouter lists the model at $0.45 per million input tokens and $3.50 per million output at the time of writing. The output price is nearly eight times the input price, and output is exactly where a thinking model spends.

Qwen recommends an output length of 32,768 tokens for most queries and 81,920 for very hard maths and programming problems. A single hard problem that uses its full 81,920-token allowance costs roughly 29 cents in output alone at that rate. That is trivial for one competition problem and expensive for a pipeline running ten thousand of them.

Two practical consequences:

• Set a per-request output cap. Pick it from measured trace lengths on your own tasks, not from the maximum. The token cost of reasoning models explains how to estimate this before you run anything at scale.
• Do not route easy work here. The model will reason about a trivial request just as it reasons about a hard one. Send direct-answer work to the Instruct 2507.

The parsing quirk that breaks integrations

The model card notes that the chat template inserts the opening <think> tag automatically. The model's output may therefore contain only the closing </think> tag, with the reasoning before it and the answer after.

Code that looks for a matched pair of tags to strip the reasoning will fail silently — usually by returning the whole trace as the answer. Split on the closing tag instead, or use a provider that separates reasoning into its own response field. Test that your provider does this rather than assuming it.

The same multi-turn rule as other Qwen3 thinking models applies: keep only final answers in the conversation history, not old reasoning traces.

Settings that matter

The card recommends temperature 0.6, top-p 0.95, top-k 20 and min-p 0. For maths, it suggests adding an instruction to reason step by step and put the final answer in a boxed expression, which also makes answers easier to extract programmatically. A presence penalty between 0 and 2 can reduce repetition, at some risk of language mixing at higher values.

When to use it, and when not to

• Use it for hard maths, algorithmic programming, and multi-step analytical problems where a wrong answer costs more than a long wait.
• Use the Instruct 2507 instead for chat, extraction, summarisation, routine coding and anything latency-sensitive.
• Think twice before using it as an agent's main model. Its TAU-bench results trailed o3 across the board, and long traces per step multiply across an agent loop.

The best test is a small one. Take thirty hard problems from your own domain with checkable answers, run them through both 2507 models, and record accuracy and output tokens per problem. If thinking flips enough wrong answers to right ones to justify the token bill, route those problem types here. How to benchmark LLMs on your own work covers the harness.