AI Progress Tracker

    DeepSeek V3 · DeepSeek-V3 (original December 2024 release)

    DeepSeek · 26 Dec 2024

    Ask the Advisor about this model
    51from 10
    benchmarks

    DeepSeek-V3 is an open-weight, text-generation model with 671 billion total parameters, approximately 37 billion activated per token, and a documented 128K-token context window. Later releases, including DeepSeek-V3.2, supersede this generation; the specifications and benchmark results here describe the original December 2024 model.

    context: 128Ktext$0.26 / $1.03 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
    AIMECompetition mathematics39.2%Dec 2024View
    GPQA DiamondPhD-level science questions59.1%Dec 2024View
    MMLU-ProBroad expert knowledge and reasoning75.9%Dec 2024View
    SWE-bench VerifiedResolving real GitHub issues42%Dec 2024View
    OTIS Mock AIME 2024-25Competition mathematics37.8%1 Apr 2025View
    FrontierMath1.7%7 Mar 2025View
    MATH Level 575.5%1 Apr 2025View
    LMArena TextHuman preference Elo from blind chat votes (not a percentage)1375 Elo22 Oct 2026View
    Aider PolyglotCode editing across six programming languages70.2%3 Oct 2025View
    LiveCodeBenchFresh competitive-programming problems49.6%1 May 2025View

    Weighted mean of 9 verified benchmark results: AIME 39.2%, GPQA Diamond 59.1%, MMLU-Pro 75.9%, SWE-bench Verified 42%, OTIS Mock AIME 2024-25 37.8%, FrontierMath 1.7%, MATH Level 5 75.5%, Aider Polyglot 70.2%, LiveCodeBench 49.6%.

    Capabilities

    • General-purpose text generation, instruction following and question answering.
    • Multilingual language processing, with published English and Chinese evaluations.
    • Code generation, code editing and software-repair assistance.
    • Mathematical and scientific problem solving, supported by published benchmark evaluations.
    • Long-document processing within a documented 128K-token context window.
    • Downloadable weights for self-hosting and infrastructure-level customization.

    What improved

    • Higher MMLU-Pro and GPQA Diamond results than DeepSeek-V2.5 in the vendor's evaluations.
    • Higher AIME 2024 and MATH-500 results than DeepSeek-V2.5.
    • Higher LiveCodeBench and SWE-bench Verified results than DeepSeek-V2.5.
    • Introduces auxiliary-loss-free expert load balancing to reduce interference between load balancing and model performance.
    • Adds a multi-token prediction training objective and uses FP8 mixed-precision training.

    Known flaws & limitations

    • The original model is text-only; native image, audio and video understanding are not documented capabilities.
    • Published mathematics, science and software-repair evaluations show substantial remaining error rates.
    • Self-hosting the complete model requires substantial memory and infrastructure despite sparse parameter activation.
    • The technical report identifies deployment efficiency and inference speed as areas requiring further improvement.

    Risks, safety & security issues

    • Incorrect answers can cause downstream harm when used for consequential decisions without independent validation.
    • Generated code and software patches require testing and security review; benchmark performance does not establish production safety.
    • A documented database exposure at DeepSeek's hosted service demonstrates the importance of protecting prompts, conversation logs and credentials; it was not an established vulnerability in V3's weights.

    Good fit for

    • Human-reviewed coding, debugging and software-maintenance assistance.
    • English and Chinese drafting, summarization and question answering.
    • Long-document analysis with independent checks of important claims.
    • Self-hosted text applications operated by teams with substantial inference infrastructure.

    Poor fit for

    • Native image, audio or video understanding.
    • Unsupervised high-stakes decisions or unattended production-code deployment.
    • Full-model deployment on memory-constrained consumer hardware.
    • Processing sensitive information through services without appropriate retention and access controls.
    Security & risk dashboardLatest vulnerabilities, advisories and real-world incidents for DeepSeek V3

    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.