Open Source Models2025-01-20
DeepSeek Releases R1 Open Reasoning Model With Benchmark Parity to o1
DeepSeek has publicly released the weights for DeepSeek-R1 and smaller distilled variants, demonstrating competitive math and coding performance through large-scale reinforcement learning.
DeepSeek has released model weights and technical documentation for DeepSeek-R1. The research demonstrates that pure reinforcement learning without preliminary supervised fine-tuning can elicit multi-step chain-of-thought reasoning behaviors. The company also published distilled models ranging from 1.5B to 70B parameter configurations on Hugging Face for local execution.
Source: DeepSeek Research AnnouncementRead Official Announcement