Qwen3 235B A22B Guide: The Original and the 2507 Instruct

"Qwen3 235B A22B" names two noticeably different models. The original, released on 29 April 2025, was a hybrid that could think or answer directly. The July update, Qwen3-235B-A22B-Instruct-2507, dropped thinking altogether, raised the native context from 32K to 256K tokens and posted much stronger numbers. Most catalogues list both, often side by side, and picking the wrong one by habit is easy.

This guide covers the shared architecture, what changed between the two, and how to choose. The thinking half of the July split has its own guide.

Reading the name: 235B of memory, 22B of compute

The A22B suffix is the important part. The model card describes a mixture-of-experts transformer with 235B total parameters and 22B activated per token: 94 layers, 128 experts of which 8 are active for any given token, and grouped-query attention with 64 query heads and 4 key-value heads.

Total parameters decide how much memory you need to hold the model. Active parameters decide how much compute each token costs. So this model needs the memory of a 235B model but runs each token at roughly the compute of a 22B dense one, which is why it can be served cheaply despite the headline size. Mixture-of-experts explained covers the mechanism, and active vs total parameters covers how to compare models that report both.

Both versions are published under Apache 2.0, so you can fine-tune, redistribute and serve them commercially without a separate agreement.

The April original: a hybrid flagship

The original shipped with the same thinking switch as the rest of Qwen3: enable_thinking in the chat template as a hard switch, and /think or /no_think in a message as a soft one. Native context is 32,768 tokens, extendable to 131,072 with YaRN.

At launch Qwen said it achieved competitive results against DeepSeek-R1, o1, o3-mini, Grok-3 and Gemini 2.5 Pro on coding, maths and general tasks, without giving the numbers in text. That comparison was against the frontier of spring 2025.

OpenRouter lists the original at a 131,072-token context and 8,192 maximum output tokens, priced at $0.455 per million input tokens and $1.82 per million output at the time of writing.

What the 2507 Instruct changed

On 21 July 2025 Qwen released the updated non-thinking version as a separate model. Per its model card it never emits thinking blocks, so there is no switch to manage, and native context rises to 262,144 tokens. Qwen later documented a configuration reaching about 1M tokens with Dual Chunk Attention, but the card puts that at roughly 1,000 GB of GPU memory, so treat it as a research option rather than a default.

The card compares it against the original in non-thinking mode, DeepSeek-V3, GPT-4o, Claude Opus 4 without thinking, and Kimi K2. Some of its figures:

• GPQA — 77.5
• AIME25 — 70.3
• LiveCodeBench v6 — 51.8
• Arena-Hard v2 — 79.2
• BFCL-v3 — 70.9, against 68.0 for the original in non-thinking mode
• MMLU-Pro — 83.0, behind Claude Opus 4 without thinking at 86.6

The fair reading is that 2507 is a large step over the original's non-thinking mode, and competitive with the strongest non-reasoning models of mid-2025 without leading everywhere. Claude Opus 4 still led on several of the tool-use and knowledge tasks in Qwen's own table.

Why the newer snapshot is cheaper

OpenRouter lists the 2507 Instruct at $0.09 per million input tokens and $0.55 per million output, with a 262,144-token context and 16,384 maximum output tokens, at the time of writing. That is cheaper than the original on both sides despite the better scores.

The architecture is identical — same parameter counts, same 22B active — so the gap is not a cheaper model to run. It reflects what the providers serving each ID charge, not anything about the weights. Prices on aggregators shift as providers enter and leave, so recheck before you build a budget on these figures.

One cost the list price hides: the original in thinking mode spends output tokens on reasoning, while the 2507 Instruct never does. For the same task, the newer model's bill can be lower by more than the price difference alone suggests.

Snapshots and pinning

The -2507 suffix is a dated snapshot: a frozen set of weights identified by its release month. Calling a dated ID means your behaviour does not change underneath you when a vendor ships an update. The undated ID is not an alias for the latest: on OpenRouter, qwen/qwen3-235b-a22b is listed as the April original, and the July model lives at qwen/qwen3-235b-a22b-2507.

For anything in production, pin the exact ID you evaluated, and treat a move to a newer snapshot as a change to test, like a dependency upgrade.

Which one to call

• Use the 2507 Instruct for almost everything that needs a direct answer: chat, extraction, tool calling, long-document work. It scores better, lists cheaper and has a far longer native context.
• Use the Thinking 2507 when the task genuinely needs extended reasoning — hard maths, competitive programming, multi-step analysis — and you can afford long outputs.
• Use the April original only where you have a validated pipeline built on its hybrid switch and the migration cost outweighs the gain.

Running both 2507 models behind a router, with the instruct as default and the thinking model as escalation, is often the best of both. Model routing and fallbacks covers how to set that up without doubling your spend.