AI Progress Tracker

    Llama 4 Scout 17B 16E Instruct · Llama-4-Scout-17B-16E-Instruct

    Meta · 5 Apr 2025

    Ask the Advisor about this model
    47from 6
    benchmarks

    Meta's Llama 4 Scout is an instruction-tuned, multimodal mixture-of-experts model with 17 billion active parameters, 16 experts and a documented 10-million-token context window. It accepts text and images and produces text, with downloadable weights governed by the Llama 4 Community License rather than a universal hosted-inference price.

    context: —text

    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 questions57.2%Apr 2025View
    MMLU-ProBroad expert knowledge and reasoning74.3%Apr 2025View
    MMMUMultimodal (image + text) understanding69.4%Apr 2025View
    OTIS Mock AIME 2024-25Competition mathematics7.8%8 Apr 2025View
    FrontierMath0%8 Apr 2025View
    MATH Level 562.3%8 Apr 2025View

    Weighted mean of 6 verified benchmark results: GPQA Diamond 57.2%, MMLU-Pro 74.3%, MMMU 69.4%, OTIS Mock AIME 2024-25 7.8%, FrontierMath 0%, MATH Level 5 62.3%.

    Capabilities

    • Processes text and image inputs together and generates text responses.
    • Supports a documented context window of up to 10 million tokens.
    • Supports multilingual assistant applications across 12 documented languages.
    • Not yet confirmed by public reporting.
    • Provides downloadable weights for customized deployment and fine-tuning, subject to its license.

    What improved

    • Extends documented context capacity from Llama 3.3's 128,000 tokens to 10 million tokens.
    • Adds native text-and-image processing compared with the text-only Llama 3.3 model.
    • Introduces a mixture-of-experts architecture in place of Llama 3's dense architecture, activating only part of the model per token.

    Known flaws & limitations

    • The documented training-data cutoff is August 2024; newer facts require external information.
    • Outputs are text only; native image, audio and video generation are not supported.
    • The advertised single-H100 deployment uses Int4 quantization, not full-precision weights.
    • Distribution and commercial use remain subject to the Llama 4 Community License and Acceptable Use Policy.

    Risks, safety & security issues

    • Incorrect or fabricated responses can mislead users, especially when consequential claims are not independently checked.
    • Biased, harmful or inappropriate outputs remain possible despite safety training.
    • Adversarial prompting can challenge model safeguards; deployed applications need additional protections and monitoring.
    • Downloadable weights permit modifications that can weaken safeguards or enable prohibited uses.

    Good fit for

    • Long-document summarization and question answering with independent checks of important claims.
    • Text-and-image assistants for visual question answering and document understanding.
    • Multilingual conversational applications in the documented supported languages.
    • Customizable, self-hosted assistants where licensing and infrastructure requirements are acceptable.

    Poor fit for

    • Autonomous medical, legal or other high-stakes decisions without qualified oversight.
    • Current-events answers without retrieval or another reliable external information source.
    • Applications requiring native audio, image or video generation.
    • Deployments requiring unrestricted licensing or assuming full-precision operation on one H100.
    Security & risk dashboardLatest vulnerabilities, advisories and real-world incidents for Llama 4 Scout 17B 16E Instruct

    References

    Profile last reviewed 10/20/2026 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.