DeepSeek R1 0528

DeepSeek released DeepSeek-R1-0528 on May 28, 2025, as a reasoning-focused thinking model with open-source weights hosted on Hugging Face. The official release announcement from DeepSeek, the model creator, lists concrete capability improvements over the prior R1 build: improved benchmark performance, enhanced front-end development capabilities, and reduced hallucinations in generated responses. These updates target reasoning quality and developer usefulness rather than core architecture changes. The same release introduced two explicit developer-facing features that were not previously available: structured JSON output and function calling support, while keeping the existing API surface unchanged. DeepSeek directed users to the Thinking Mode guide and API docs for integration details. The release is historical context for the deepseek-thinking family, preserved here with creator attribution to DeepSeek rather than the Alibaba DashScope hosting layer.

DeepSeek released DeepSeek-R1-0528 as a minor version upgrade to the original R1, announced via the official DeepSeek API documentation on May 28, 2025. The release highlights improved benchmark performance, enhanced front-end capabilities, reduced hallucinations, and new support for JSON output and function calling. API usage was unchanged from the prior R1 release. Open-source weights for DeepSeek-R1-0528 were published on Hugging Face, making the updated reasoning model freely available. The release positions R1-0528 as a reasoning-focused model whose capabilities were broadened with structured output and tool-calling features, while keeping the same developer-facing API surface as before.

A third-party deep dive analyzes DeepSeek-R1-0528 as more than a cosmetic patch, detailing that it retains the Mixture-of-Experts architecture scaled up with 128K context (extendable via RoPE scaling) while delivering large mathematical-reasoning improvements. The post highlights the AIME 2025 jump from 70% to 87.5% as The piece contextualizes R1-0528 within DeepSeek's open-source trajectory and notes improvements relevant to vibe coding and function calling. It frames the release as a stealth incremental update that nonetheless shifts the competitive balance against closed frontier models in mathematical and multi-step reasoning wor

A Medium post by Joe Njenga (May 29, 2025) documents the quiet release of DeepSeek R1-0528 to Hugging Face without a formal announcement. The piece confirms the model's architecture: 671 billion total parameters with 37 billion active during inference, achieving large-model performance with reduced computational overhead. It positions R1-0528 as gaining rapid traction in the open-source community shortly after release. The article frames R1-0528 as a community-discovered reasoning model update rather than a full R2 successor, noting benchmark performance rivaling OpenAI's o3. It provides historical release context from launch week, documenting how developers found and tested the model within hours of its Hugging Face upload. The body is paywalled beyond these visible excerpts.

BentoML's developer guide (last updated April 24, 2026) includes a dedicated DeepSeek-R1-0528 section within its broader DeepSeek model lineup overview spanning V3, R1, V3.1, V3.2, V4, and distilled variants. The guide positions R1-0528 as a refined reasoning model alongside its predecessor R1, helping developers choose among variants. It frames the model within DeepSeek's rapid lineup expansion and the 671B-parameter architecture pattern shared with V3.1. The BentoML overview situates R1-0528 among DeepSeek's reasoning-focused releases, contrasting it with base and chat variants of V3 and later V3.1/V4 alternatives. It highlights practical deployment considerations and clarifies naming confusion across R1, R1-Zero, and the 0528 update. The guide serves as independent technical context for developers evaluating where R1-0528 fits in the DeepSeek ecosystem.

The official DeepSeek-R1-0528 model card on Hugging Face documents a major reasoning upgrade over the prior R1, driven by increased post-training compute and algorithmic optimization. It reports large benchmark gains, including AIME 2025 accuracy rising from 70% to 87.5% and GPQA-Diamond Pass@1 climbing from 71.5 to 81.0, alongside higher average token usage per AIME question. Beyond reasoning improvements, the model card notes a reduced hallucination rate, enhanced function calling support, and a better vibe-coding experience. Generation length is capped at 64K tokens, and the model was evaluated at temperature 0.6 with top-p 0.95 and 16 samples for pass@1 estimation across mathematics, programming, and general logic benchmarks.