AI Progress Tracker

    Qwen3 32B · Qwen3-32B

    Alibaba · 29 Apr 2025

    Ask the Advisor about this model
    57from 4
    benchmarks

    Qwen3-32B is Alibaba's open-weight, text-only language model with switchable thinking and non-thinking modes, multilingual support, and tool-use capabilities. Newer Qwen3-2507 family checkpoints have shipped, but this profile covers the original 32-billion-parameter checkpoint.

    context: 33Ktext$0.08 / $0.28 per 1M tokens (in / out)

    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 questions65.7%28 Aug 2026View
    OTIS Mock AIME 2024-25Competition mathematics66.9%30 Aug 2026View
    LMArena TextHuman preference Elo from blind chat votes (not a percentage)1340 Elo13 Sep 2026View
    Aider PolyglotCode editing across six programming languages40%8 May 2025View

    Weighted mean of 3 verified benchmark results: GPQA Diamond 65.7%, OTIS Mock AIME 2024-25 66.9%, Aider Polyglot 40%.

    Capabilities

    • Switches between thinking mode for complex problems and non-thinking mode for faster responses.
    • Supports 119 languages and dialects.
    • Handles mathematical reasoning, programming, and general text-based assistance.
    • Supports tool calling and integration with Qwen-Agent.
    • Provides open weights under the Apache 2.0 license for self-hosting and customization.

    What improved

    • Combines thinking and non-thinking behavior in one checkpoint, unlike standard Qwen2.5 instruction models.
    • Expands multilingual coverage from Qwen2.5's 29 languages to 119 languages and dialects.
    • Improves reasoning, mathematics, and coding performance in Alibaba's release evaluations.
    • Adds dedicated post-training for stronger tool-use and agent capabilities.

    Known flaws & limitations

    • The 131,072-token context requires YaRN extension; the native context limit is 32,768 tokens.
    • Accepts text rather than native image, audio, or video inputs.
    • Thinking mode generates additional reasoning tokens, increasing latency and inference consumption.
    • Static YaRN scaling can reduce performance on shorter inputs.

    Risks, safety & security issues

    • Not yet confirmed by public reporting.

    Good fit for

    • Self-hosted multilingual text assistants.
    • Coding assistance and mathematical problem-solving with review of generated answers.
    • Tool-enabled assistants using Qwen-Agent or compatible frameworks.
    • Applications that need selectable reasoning depth and response latency.

    Poor fit for

    • Native image, audio, or video understanding.
    • Long-context workloads beyond 32,768 tokens without a configured and tested context-extension setup.
    • Strict low-latency workloads that require thinking mode on every request.
    Security & risk dashboardLatest vulnerabilities, advisories and real-world incidents for Qwen3 32B

    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.