AI Progress Tracker

    Qwen3 235B A22B · qwen3-235b-a22b

    Alibaba · 29 Apr 2025

    Ask the Advisor about this model
    70from 5
    benchmarks

    Alibaba's Qwen3-235B-A22B is an open-weight, text-only mixture-of-experts model with 235 billion total parameters and 22 billion activated per token, supporting both thinking and non-thinking modes. It remains the original hybrid-thinking release, distinct from the separate Instruct-2507 and Thinking-2507 successors released in July 2025.

    context: 33Ktext

    Benchmark results

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

    BenchmarkScoreSource
    GPQA DiamondPhD-level science questions70.7%3 Jun 2025View
    MATH Level 568.9%3 Jun 2025View
    LMArena TextHuman preference Elo from blind chat votes (not a percentage)1366 Elo22 Oct 2026View
    Aider PolyglotCode editing across six programming languages59.6%View
    LiveCodeBenchFresh competitive-programming problems80.4%1 May 2025View

    Weighted mean of 4 verified benchmark results: GPQA Diamond 70.7%, MATH Level 5 68.9%, Aider Polyglot 59.6%, LiveCodeBench 80.4%.

    Capabilities

    • Switches between thinking and non-thinking modes within one model.
    • Supports 119 languages and dialects.
    • Provides mathematical, scientific and coding reasoning in thinking mode.
    • Supports tool calling and agent workflows, including integration through Qwen-Agent.
    • Offers Apache 2.0-licensed open weights for self-hosting and customization.

    What improved

    • Combines deliberate reasoning and direct responses in one model, rather than Qwen2.5's instruction-focused approach.
    • Expands language coverage from Qwen2.5's 29 languages to 119 languages and dialects.
    • Uses approximately 36 trillion pretraining tokens, compared with approximately 18 trillion for Qwen2.5.
    • The vendor reports improvements over Qwen2.5 in reasoning, coding, instruction following and tool use.

    Known flaws & limitations

    • The native context window is 32,768 tokens; reaching 131,072 requires YaRN configuration.
    • Text-only inputs and outputs do not provide native image, audio or video understanding.
    • The 235-billion-parameter weight set requires substantial deployment resources despite activating only 22 billion parameters per token.
    • The model card warns that greedy decoding can degrade performance and cause endless repetitions.
    • Static YaRN scaling can affect performance on shorter inputs.

    Risks, safety & security issues

    • Not yet confirmed by public reporting.

    Good fit for

    • Self-hosted multilingual text assistants where substantial inference infrastructure is available.
    • Mathematical and scientific problem-solving with independently checked outputs.
    • Coding assistance and code-editing workflows with tests and review.
    • Tool-using assistants that need selectable thinking and non-thinking behavior.

    Poor fit for

    • Applications requiring native image, audio or video understanding.
    • On-device or memory-constrained deployments.
    • Long-document workloads exceeding the native context window without context-extension configuration and validation.
    • Deployments relying on greedy decoding for reliable output.
    Security & risk dashboardLatest vulnerabilities, advisories and real-world incidents for Qwen3 235B A22B

    References

    Profile last reviewed 10/27/2026 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.