π¬ Distilled from DeepSeek-R1, 6 small models fully open-sourced
π 32B & 70B models on par with OpenAI-o1-mini
π€ Empowering the open-source community
π Pushing the boundaries of **open AI**!
π 2/n
π License Update!
π DeepSeek-R1 is now MIT licensed for clear open access
π Open for the community to leverage model weights & outputs
π οΈ API outputs can now be used for fine-tuning & distillation
π 3/n
π οΈ DeepSeek-R1: Technical Highlights
π Large-scale RL in post-training
π Significant performance boost with minimal labeled data
π’ Math, code, and reasoning tasks on par with OpenAI-o1
π More details:
π Exciting news! Weβve officially launched DeepSeek-V2.5 β a powerful combination of DeepSeek-V2-0628 and DeepSeek-Coder-V2-0724! Now, with enhanced writing, instruction-following, and human preference alignment, itβs available on Web and API. Enjoy seamless Function Calling, FIM, and Json Output all-in-one!
Note: Due to significant updates in this version, if performance drops in certain cases, we recommend adjusting the system prompt and temperature settings for the best results!
DeepSeek-V2.5 outperforms both DeepSeek-V2-0628 and DeepSeek-Coder-V2-0724 on most benchmarks.
In our internal Chinese evaluations, DeepSeek-V2.5 shows a significant improvement in win rates against GPT-4o mini and ChatGPT-4o-latest (judged by GPT-4o) compared to DeepSeek-V2-0628.