StorySoftware
DeepSeek releases R1 open-weight reasoning model under MIT licence
DeepSeek released R1, a reasoning model achieving performance on par with OpenAI's o1, with weights under the MIT licence and six smaller distilled models. Its paper showed reasoning could be taught largely through reinforcement learning rather than human-written examples; API prices were a small fraction of o1's.
- Open-weight under MIT license; performs on par with OpenAI o1
- Six smaller distilled models available from 32B to 70B parameters; smaller versions match o1-mini
- Reasoning taught via large-scale RL rather than human examples; novel training approach
- API pricing: $0.14/M input (cache hit), $0.55/M (cache miss), $2.19/M output—fraction of o1 pricing