DeepSeek quietly drops an updated R1 on Hugging Face

DeepSeek-R1-0528 is the latest minor refresh of DeepSeek's flagship R1 reasoning model. It sharpens the model's skills across mathematics, coding, and complex logic, pushing overall performance close to frontier closed models like O3 and Gemini 2.5 Pro.

What it can do

This update focuses entirely on post-training improvements. DeepSeek poured more compute into this phase and introduced new algorithmic optimization techniques, encouraging the model to think significantly deeper before answering.

That deeper chain of thought shows up clearly on difficult math benchmarks. On AIME 2025, accuracy jumped from 70% in the original R1 to 87.5%. To pull that off, the new version spends an average of 23K thinking tokens per question, up from 12K previously.

Other benchmarks see similar jumps. GPQA-Diamond pass rates climbed from 71.5% to 81.0%, while MMLU-Redux hit 93.4% and MMLU-Pro reached 85.0%. Outside of raw benchmarks, DeepSeek reports fewer hallucinations, cleaner function calling, and a noticeably better flow for vibe coding workflows. The maximum generation length across the board is set at 64K tokens.

How to run it and hardware specs

The full model weights and configuration files are available right now on Hugging Face under a permissive MIT license, meaning commercial deployments are fully welcome.

Running it locally is another story. Weighing in at 685 billion parameters, this is a massive model. Without heavy quantization or architectural trimming, it will not run on consumer-grade PCs or single enthusiast GPUs. You will need high-end multi-GPU server infrastructure to host it.

If you have the hardware, deployment is turnkey. The Hugging Face repository provides ready-to-use launch commands for vLLM, SGLang, Hugging Face Transformers, and Docker Model Runner. Serving an OpenAI-compatible API locally through vLLM or SGLang takes just a single command.

Why it matters

Most developers will not be downloading hundreds of gigabytes of weights to their laptops, but having the unredacted 685B model open source means inference providers will have cheap hosted endpoints up almost immediately. You get near-O3 reasoning performance without being locked into proprietary APIs.

It also proves that post-training compute scaling is nowhere near hitting a wall. By letting the model chew on 23K tokens of internal reasoning, DeepSeek squeezed massive benchmark gains out of an existing base. If your workflow relies on heavy coding tasks, complex math, or automated tool chains, switching to the 0528 weights is well worth an immediate run.