AI Progress Tracker

    Qwen3 235B A22B Thinking · Qwen3-235B-A22B-Thinking-2507

    Alibaba · 25 Jul 2025

    Ask the Advisor about this model
    64from 8
    benchmarks

    Alibaba's Qwen3-235B-A22B-Thinking-2507 is an open-weight, text-only reasoning model with 235 billion total parameters and 22 billion activated per token. This thinking-only checkpoint supports a native 262,144-token context and publishes vendor evaluations covering mathematics, scientific reasoning, coding and general knowledge.

    context: 262Ktext

    Benchmark results

    Every figure links to where it was published. Compare like with like — harnesses and prompt setups differ between vendors.

    BenchmarkScoreSource
    AIMECompetition mathematics92.3%Jul 2025View
    GPQA DiamondPhD-level science questions81.1%Jul 2025View
    Humanity's Last ExamFrontier expert questions across fields18.2%Jul 2025View
    LiveCodeBenchFresh competitive-programming problems74.1%Jul 2025View
    MMLU-ProBroad expert knowledge and reasoning84.4%Jul 2025View
    OTIS Mock AIME 2024-25Competition mathematics86.7%10 Dec 2025View
    FrontierMath8.5%11 Dec 2025View
    SimpleQA VerifiedShort factual questions (hallucination resistance)40.4%27 Aug 2026View

    Weighted mean of 8 verified benchmark results: AIME 92.3%, GPQA Diamond 81.1%, Humanity's Last Exam 18.2%, LiveCodeBench 74.1%, MMLU-Pro 84.4%, OTIS Mock AIME 2024-25 86.7%, FrontierMath 8.5%, SimpleQA Verified 40.4%.

    Capabilities

    • Multi-step mathematical reasoning, supported by published competition-mathematics evaluations.
    • Scientific question answering, evaluated on GPQA Diamond.
    • Code generation and algorithmic problem solving, evaluated on LiveCodeBench v6.
    • Long-document processing within a native 262,144-token context window.
    • Open-weight deployment with a mixture-of-experts architecture that activates 22 billion parameters per token.

    What improved

    • Native context expands from the original checkpoint's 32,768 tokens to 262,144 tokens.
    • Vendor evaluations report stronger mathematical and scientific reasoning than the original Qwen3-235B-A22B.
    • Vendor evaluations report improved coding performance over the original checkpoint.
    • A dedicated thinking-only configuration replaces the predecessor's switchable thinking and non-thinking modes.

    Known flaws & limitations

    • Text-only: it does not natively process images, audio or video.
    • This checkpoint does not support switching to a non-thinking mode.
    • Self-hosting must accommodate 235 billion total parameters despite the smaller active parameter count.
    • Its published Humanity's Last Exam result shows substantial remaining limitations on exceptionally difficult expert questions.

    Risks, safety & security issues

    • Not yet confirmed by public reporting.

    Good fit for

    • Mathematical and scientific problem solving with independently checked answers.
    • Programming assistance and algorithmic tasks backed by executable tests.
    • Long-document analysis and synthesis using text inputs.
    • Self-hosted reasoning applications with substantial inference infrastructure.

    Poor fit for

    • Native image, audio or video understanding.
    • Applications requiring a selectable non-thinking mode.
    • Resource-constrained local deployment.
    • Unsupervised high-stakes decisions requiring consistently correct expert answers.
    Security & risk dashboardLatest vulnerabilities, advisories and real-world incidents for Qwen3 235B A22B Thinking

    References

    Profile last reviewed 10/26/2026 and re-checked automatically every few hours against the vendor's published documentation, papers and public leaderboards. Scores come from the cited benchmark results above — confirm limits and pricing against the provider's own documentation before building on them.