AI Progress Tracker

    Deepseek R1 0528 Qwen3 8b · DeepSeek-R1-0528-Qwen3-8B

    28 May 2025

    Ask the Advisor about this model
    64from 4
    benchmarks

    DeepSeek-R1-0528-Qwen3-8B is an open-weight, text-only reasoning model built on Qwen3-8B using reasoning data from DeepSeek-R1-0528. Its published evaluations demonstrate mathematics, science and coding capabilities, while its compact size supports self-hosted deployment.

    context: 131Ktext

    Benchmark results

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

    BenchmarkScoreSource
    AIMECompetition mathematics86%May 2025View
    GPQA DiamondPhD-level science questions61.1%May 2025View
    LiveCodeBenchFresh competitive-programming problems60.5%May 2025View
    OTIS Mock AIME 2024-25Competition mathematics43.9%28 Aug 2026View

    Weighted mean of 4 verified benchmark results: AIME 86%, GPQA Diamond 61.1%, LiveCodeBench 60.5%, OTIS Mock AIME 2024-25 43.9%.

    Capabilities

    • Solves competition-style mathematics problems using extended reasoning.
    • Answers challenging scientific questions, with published GPQA Diamond results.
    • Generates solutions to programming problems, with published LiveCodeBench results.
    • Provides downloadable weights for local deployment and customization.
    • Supports deployment through inference frameworks including vLLM and SGLang.

    What improved

    • Uses Qwen3-8B rather than the Qwen2.5-Math backbone of DeepSeek-R1-Distill-Qwen-7B.
    • Learns from reasoning data generated by the updated DeepSeek-R1-0528 teacher.
    • Reports higher AIME 2024 performance than DeepSeek-R1-Distill-Qwen-7B.
    • Reports higher GPQA Diamond and LiveCodeBench performance than DeepSeek-R1-Distill-Qwen-7B.

    Known flaws & limitations

    • Text-only model; it does not natively process images, audio or video.
    • Published mathematics, science and coding evaluations show that answers are not consistently correct.
    • DeepSeek warns that greedy decoding can cause endless repetition.
    • The model can bypass its thinking pattern; DeepSeek recommends explicitly starting generation with the thinking prefix.

    Risks, safety & security issues

    • Incorrect scientific answers or generated code can cause downstream harm when used without independent validation.
    • Repetitive or lengthy reasoning can consume substantial inference resources; deployments should enforce token and execution limits.
    • Hosted deployments can expose sensitive conversations if logging databases and credentials are inadequately protected; this is a deployment risk, not an established checkpoint vulnerability.

    Good fit for

    • Self-hosted mathematics and reasoning assistance with answer checking.
    • Programming assistance backed by automated tests and code review.
    • Research into reasoning distillation and compact language models.
    • Text-based scientific question answering with expert verification.

    Poor fit for

    • Image, audio or video understanding without separate modality-processing components.
    • Unsupervised medical, legal or other high-stakes decision-making.
    • Executing generated code directly in privileged production environments.
    • Applications requiring consistently short, tightly bounded response latency.
    Security & risk dashboardLatest vulnerabilities, advisories and real-world incidents for Deepseek R1 0528 Qwen3 8b

    References

    Profile last reviewed 10/21/2026 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.