Open R1: Update #4
Hugging Face
DeepSeek quietly dropped an updated V3 model, now MIT-licensed. It's beating Claude-3.7-Sonnet on math and coding benchmarks and it's free to fork.
Based on reporting by Hugging Face — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
DeepSeek didn't bother with a launch event. The company just pushed an updated version of its V3 base model onto Hugging Face this week and let people find it. No press release, no fanfare — just a quiet drop of DeepSeek-V3-0324, and it turns out to be a meaningfully better model than what came before.
The architecture hasn't changed. What has changed is the license, now MIT instead of the custom terms DeepSeek used previously, which matters a lot for anyone building commercial products on top of it. The gains are real too: MMLU-Pro jumped from 75.9 to 81.2, AIME math scores climbed nearly 20 points to 59.4, and LiveCodeBench rose 10 points to 49.2. Hugging Face's own comparison puts it roughly level with GPT-4.5 and ahead of Claude-3.7-Sonnet on several of these tasks, which is not a small claim for an open-weight model.
DeepSeek's model card points to specific fixes rather than vague improvements: better front-end code that actually runs and looks decent, sharper Chinese writing that matches the R1 house style, more reliable function calling, and better multi-turn rewriting. No technical report has landed yet, so nobody outside the company knows exactly what changed under the hood. Hugging Face's best guess is a mix of continual pretraining on fresher, cleaner data plus a reworked post-training recipe, since that's usually where instruction-following and style gains come from these days.
Getting the model running is already straightforward. It's live through Hugging Face's Inference Providers via Fireworks, Hyperbolic and Novita, and it works out of the box with TGI and SGLang on H100 clusters. For anyone without a spare node of H100s lying around, Unsloth's dynamic quantization cuts the compute requirement roughly in half with only a small hit to benchmark scores, running through llama.cpp.
The safety section of the post is really a broader statement about open models, not just DeepSeek. Hugging Face argues that downloading and running weights in safetensors format carries no hidden code-execution risk, and that the real exposure comes later — in trusting code the model writes without review, or letting an agent run unsandboxed with access to real credentials. That's a fair distinction, and one that gets lost whenever
My take — AI-written commentary, not fact-checked reporting
, but it's easy to conflate them. Editor's note: trimming trailing artifact.
Read more about this at: Hugging Face
