qwen3-235b-a22b-instruct-2507

Overview. Qwen3-235B-A22B-Instruct-2507 is the July 2025 instruction-tuned release of Qwen's largest Qwen3 model. It is a Mixture-of-Experts model with 235 billion parameters in total. It activates 22 billion of them per token (8 of 128 experts), so it runs at much lower cost than its total size suggests. It handles text in and text out, accepts up to 262,144 tokens of context, and is released under the Apache 2.0 licence.

Key strengths. - Strong multilingual performance, among the higher scores in its class - Large context window for long documents and retrieval - Low cost for its size, thanks to the sparse MoE design - Open weights under Apache 2.0, so the same model can be served by several providers

When to use it. Use it when you need a capable general-purpose model at a low price per token. It suits long documents, multilingual work, summarisation, extraction and chat assistants where a direct answer is better than a visible chain of thought. It also fits workloads that need predictable latency. It returns no reasoning tokens, so output length and cost stay close to the answer itself.

Limitations. This is the non-thinking variant, and it ignores reasoning-effort settings. For multi-step maths, planning or hard coding problems, a reasoning model will usually do better. Qwen publishes a separate Thinking-2507 variant for those tasks. The model accepts text only, with no image or audio input, and returns at most 16,384 tokens per response.

Sovereignty note. The model is developed by Qwen, part of Alibaba Group, and published as open weights. The weights can be run anywhere, so the creator never sees your prompts. Where your data is processed, and under which retention and training terms, depends on the endpoint that serves the request, not on the model.