DeepSeek-R1-0528

What it is

DeepSeek-R1-0528 is an updated version of DeepSeek's reasoning model DeepSeek-R1, released on 28 May 2025.^[1][2] It works the way R1 does: the model writes out a chain of thought before it commits to an answer, and that visible reasoning is what makes it useful on hard problems. The release is published under the MIT License, which permits commercial use and also permits distilling the model into other models.^[1]

What the benchmarks show

The update is mostly a story of numbers moving up. On AIME 2025, a competition-mathematics test, accuracy rose from 70.0% to 87.5%.^[1] The gain came with more thinking rather than faster thinking: the average number of tokens the model spent on an AIME question grew from about 12K to about 23K, so it worked through roughly twice as much reasoning before answering.^[1] The same pattern holds on other math tests — AIME 2024 went from 79.8 to 91.4, and HMMT 2025 from 41.7 to 79.4, according to the model card's comparison with DeepSeek R1.^[1]

Science and coding moved as well. GPQA-Diamond, a set of hard graduate-level science questions, rose from 71.5 to 81.0, and LiveCodeBench (2408–2505) from 63.5 to 73.3.^[1] On software engineering, SWE-bench Verified — which measures whether a model can resolve real issues in code repositories — climbed from 49.2 to 57.6.^[1] Humanity's Last Exam, a broad and deliberately difficult expert question set, went from 8.5 to 17.7.^[1] Scores on that last benchmark are low in absolute terms even after the update, which is a reminder that these tests are designed to leave plenty of headroom.

Behaviour and interface changes

Beyond scores, DeepSeek described two behavioural changes: a reduced hallucination rate, and a better experience for what it calls vibe coding, the practice of building software by describing what you want and letting a model write the code.^[1] The interface changed too. The model now supports system prompts, and users no longer have to force a `<think>` prefix into the input to trigger reasoning — the model reasons when it should rather than when prompted with a marker.^[1] Function calling was enhanced as well, with a BFCL v3 multi-turn score of 37.0 reported.^[1] Taken together, these are the changes that matter to people building applications on top of the model rather than to people reading leaderboards.

The distilled sibling

DeepSeek also distilled the model's chain of thought into Qwen3 8B Base, producing DeepSeek-R1-0528-Qwen3-8B.^[1] That small model scored 86.0 on AIME 2024 — 10 points above Qwen3 8B and matching Qwen3-235B-thinking, a model many times its size.^[1] Since the license allows distillation, this release doubles as a demonstration of how far reasoning ability can be carried into a much smaller model.

Size and serving

The Hugging Face model card lists DeepSeek-R1-0528 at 685B parameters, stored as BF16, F8_E4M3 and F32 tensors, and sets the maximum generation length at 64K tokens.^[1] For deployment it recommends a temperature of 0.6 in the web and app environments, and notes that the model can be served with vLLM or SGLang behind an OpenAI-compatible chat completions endpoint.^[1] In other words, it is a large model intended to be run as a service rather than on a laptop.

Where it sits

DeepSeek-R1-0528 was itself superseded: DeepSeek-V3.1 arrived on 21 August 2025, and its thinking mode scored 88.4 on AIME 2025 while answering faster than R1-0528.^[3] That short gap between releases is a fair summary of the period — an open-weight reasoning model, improved in place, then folded into the next generation within a few months.