AI Progress Tracker

    Qwen3 14B · qwen3-14b

    Alibaba · 29 Apr 2025

    Ask the Advisor about this model
    65from 2
    benchmarks

    Qwen3-14B is Alibaba's open-weight, text-only language model with integrated thinking and non-thinking modes, multilingual support, and tool-use capabilities. Newer Qwen3-2507 family releases are available, but this model retains a native 32,768-token context window, extendable to 131,072 tokens with YaRN.

    context: 33Ktext$0.12 / $0.24 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 questions63.8%28 Aug 2026View
    OTIS Mock AIME 2024-25Competition mathematics66.4%28 Aug 2026View

    Weighted mean of 2 verified benchmark results: GPQA Diamond 63.8%, OTIS Mock AIME 2024-25 66.4%.

    Capabilities

    • Switches between thinking and non-thinking modes within a single model.
    • Supports 119 languages and dialects, according to the Qwen3 release documentation.
    • Supports mathematical reasoning, code generation, and general text-based assistance.
    • Supports tool calling and integration with the Qwen-Agent framework.
    • Provides downloadable weights under the Apache 2.0 license for local deployment and customization.

    What improved

    • Compared with Qwen2.5-14B-Instruct, integrates thinking and non-thinking behavior into one model.
    • Adds explicit reasoning-mode controls, allowing applications to balance reasoning effort and response latency.
    • Expands documented multilingual coverage from the Qwen2.5 generation's 29 languages to 119 languages and dialects.
    • The Qwen3 release documents family-level improvements in reasoning, instruction following, and agent-oriented tool use.

    Known flaws & limitations

    • Text-only: image, audio, and video understanding are not native capabilities.
    • The advertised 131,072-token maximum requires YaRN configuration; the native context window is 32,768 tokens.
    • Thinking mode generates additional reasoning tokens, increasing generation time and resource consumption.
    • The model card advises against greedy decoding because it can degrade performance and cause endless repetition.
    • Static YaRN scaling can affect performance on shorter inputs; the model card recommends enabling it only when long-context processing is required.

    Risks, safety & security issues

    • Not yet confirmed by public reporting.

    Good fit for

    • Locally hosted text assistants requiring downloadable, permissively licensed weights.
    • Multilingual conversation and text-generation applications.
    • Coding and mathematical assistance with configurable reasoning effort.
    • Developer-controlled tool-using assistants built with an agent framework.

    Poor fit for

    • Applications requiring native image, audio, or video understanding.
    • Processing more than 32,768 tokens without configuring context extension.
    • Strict low-latency applications that leave thinking mode enabled.
    • Deployments that require greedy decoding without repetition-related safeguards.
    Security & risk dashboardLatest vulnerabilities, advisories and real-world incidents for Qwen3 14B

    References

    Profile last reviewed 10/18/2026 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.