DeepSeek updates its R1 reasoning AI model and releases it on Hugging Face

DeepSeek's R1 AI model, now 685 billion parameters big, expands to Hugging Face for commercial use.

: DeepSeek, a Chinese startup, has updated its R1 reasoning AI model and launched it on the developer platform Hugging Face, making it commercially available under the permissive MIT license. The model weighs in at 685 billion parameters, creating a challenge for running on consumer hardware. DeepSeek gained attention by rivaling OpenAI's models and faced regulatory scrutiny over national security concerns. The repository on Hugging Face includes only configuration files and weights, lacking a detailed description.

DeepSeek, a Chinese AI startup, has released an updated version of its R1 reasoning model, named R1-0528, on Hugging Face. This release was done quietly via a WeChat post, without an official press release. Despite the soft announcement, the model has shown strong results in code generation benchmarks on LiveCodeBench, trailing just behind OpenAI’s o4 mini and o3, while outperforming competitors like xAI’s Grok 3 mini and Alibaba’s Qwen 3.

The original DeepSeek-R1 model, launched in January 2025, gained attention for delivering performance comparable to OpenAI’s o1 model in math, coding, and reasoning—at a significantly lower cost. The model was developed using a novel training approach: reinforcement learning (RL) without supervised fine-tuning (SFT), along with pre-RL cold-start data. Its backbone is based on DeepSeek-V3-Base and it supports a context window of up to 128K tokens.

To support the AI research community, DeepSeek has open-sourced not only the main R1 model but also multiple distilled variants built on Qwen and Llama architectures. These models range from 1.5B to 70B parameters and have posted strong benchmark results. Notably, DeepSeek-R1-Distill-Qwen-32B outperforms OpenAI’s o1-mini in several reasoning and code-generation tasks.

DeepSeek’s efficient development process has challenged the dominant narrative that top-tier AI models require massive resources. The company claims its V3 model cost under $6 million to train—substantially lower than OpenAI’s estimated $100 million expenditure for GPT-4. This cost-effective approach is shaking up the competitive dynamics of the AI industry, with other firms adjusting pricing and release schedules in response.

While the AI community awaits the expected release of DeepSeek’s more advanced R2 model, the R1-0528 update continues to reflect the company’s mission of democratizing access to powerful reasoning models. All versions of DeepSeek-R1 are freely available on Hugging Face, inviting developers and researchers worldwide to experiment and build with open-source AI.

Sources: Reuters, TechCrunch, Hugging Face

TechnologyAIstartups