Qwen3-2507 (Qwen3-235B-A22B Instruct and Thinking)

2025-07-21 · Publisher : Alibaba · Domain : Artificial Intelligence

Alibaba Qwen releases Qwen3-235B-A22B-Instruct-2507 (July 21) then Qwen3-235B-A22B-Thinking-2507 (July 25): two updates of the 235B MoE, with a native context of 262,144 tokens, under Apache 2.0.

Alibaba Qwen releases Qwen3-235B-A22B-Instruct-2507 (July 21) then Qwen3-235B-A22B-Thinking-2507 (July 25): two updates of the 235B MoE, with a native context of 262,144 tokens, under Apache 2.0. ### Context Three months after Qwen3, Qwen drops the hybrid mode in favor of two specialized models, one for direct answers, the other for long reasoning. The comparison is then made against Claude Opus 4, GPT-4o, DeepSeek-V3-0324, Kimi K2, OpenAI o3 and o4-mini, Gemini 2.5 Pro and DeepSeek-R1-0528. ### What the model brings - Instruct-2507 supports only the direct mode and improves in instruction following, logic, mathematics, science, code and tools. - Thinking-2507 supports only the thinking mode, with increased thinking length; the vendor describes it as state of the art among open thinking models. - Native context of 262,144 tokens, extensible up to 1,010,000 tokens (update of August 8, 2025 according to the GitHub repository). - Coverage of long-tail knowledge in several languages and stronger alignment on subjective tasks. - The 2507 lineup was later extended to the 30B-A3B (July 30 and 31) and the 4B (August 6). ### Architecture and specifications MoE with 235 billion total parameters, 22 billion active, 94 layers, 128 experts with 8 active, GQA attention with 64 query heads for 4 key-value heads. Apache 2.0 license. Knowledge cutoff: not disclosed. Both models were put online on Hugging Face on July 21 and 25, 2025. ### Cost and access | Item | Value | |---|---| | Weights | Free, Apache 2.0 | | Alibaba Cloud API Instruct-2507 (current pricing consulted) | $0.23 input / $0.92 output per million tokens | | Alibaba Cloud API Thinking-2507 (current pricing consulted) | $0.23 input / $2.3 output per million tokens | ### Benchmarks Instruct-2507, Hugging Face model card (vendor measurements). | Benchmark | GPT-4o-0327 | Claude Opus 4 (without thinking) | Qwen3-235B-A22B (former direct mode) | Instruct-2507 | |---|---|---|---|---| | GPQA | 66.9 | 74.9 | 62.9 | 77.5 | | AIME25 | 26.7 | 33.9 | 24.7 | 70.3 | | LiveCodeBench v6 | 35.8 | 44.6 | 32.9 | 51.8 | | Arena-Hard v2 | 61.9 | 51.5 | 52.0 | 79.2 | | MMLU-Pro | 79.8 | 86.6 | 75.2 | 83.0 | Thinking-2507, Hugging Face model card (vendor measurements). | Benchmark | OpenAI o3 | Gemini 2.5 Pro | Claude Opus 4 (thinking) | Thinking-2507 | |---|---|---|---|---| | GPQA | 83.3 | 86.4 | 79.6 | 81.1 | | AIME25 | 88.9 | 88.0 | 75.5 | 92.3 | | HMMT25 | 77.5 | 82.5 | 58.3 | 83.9 | | LiveCodeBench v6 | 58.6 | 72.5 | 48.9 | 74.1 | | Arena-Hard v2 | 80.8 | 72.5 | 59.1 | 79.7 | ### Versus closed models In direct mode, Instruct-2507 outperforms GPT-4o-0327 on every row of the table and outperforms Claude Opus 4 without thinking on GPQA (77.5 versus 74.9), AIME25 and LiveCodeBench v6, but remains behind on MMLU-Pro (83.0 versus 86.6). In thinking mode, the open model is ahead of o3 and Gemini 2.5 Pro on AIME25 (92.3 versus 88.9 and 88.0) and on LiveCodeBench v6 (74.1 versus 72.5 for Gemini), but behind on GPQA (81.1 versus 86.4 for Gemini). These figures are published by the vendor with its own protocols; the model card notes that some competitor scores come from its own evaluations. Alibaba's API pricing is given above; no sourced price comparison with closed APIs is established here. ### Key takeaways Qwen3-2507 shows the gain obtained by specializing each mode. The 235B MoE becomes an open model at the level of the best closed models of summer 2025 on mathematical reasoning and code.

Tags : LLM, Open Source, Alibaba, Qwen, MoE, Apache 2.0, Raisonnement