DeepSeek-R1 is a reinforcement learning (RL) driven inference model that addresses issues of repetitiveness and readability within the model. Prior to RL, DeepSeek-R1 introduced cold start data to further optimize inference performance. It performs comparably to OpenAI-o1 in mathematical, coding, and reasoning tasks, and enhances overall effectiveness through meticulously designed training methods.

DeepSeek-V3 is a MoE model developed by Hangzhou DeepSeek Artificial Intelligence Technology Research Co., Ltd., achieving outstanding results in multiple evaluations and ranking first among open-source models on mainstream leaderboards. Compared to the V2.5 model, V3 has achieved a threefold increase in generation speed, providing users with a faster and smoother experience.

The QwQ inference model is trained based on the Qwen2.5-32B model, significantly enhancing its reasoning capabilities through reinforcement learning. The core metrics of the model, including mathematical code (AIME 24/25, LiveCodeBench) and some general metrics (IFEval, LiveBench, etc.), reach the level of the full version of DeepSeek-R1, with all metrics significantly surpassing those of DeepSeek-R1-Distill-Qwen-32B, which is also based on Qwen2.5-32B.

The DeepSeek-R1-Distill series models are fine-tuned versions of samples generated by DeepSeek-R1, using knowledge distillation techniques on open-source models like Qwen and Llama.

qwen2.5-72b-instruct.description

qwen2.5-32b-instruct.description

qwen2.5-coder-32b-instruct.description

qwen2.5-14b-instruct.description

qwen2.5-7b-instruct.description

Qwen2 is the new generation of large language model series introduced by the Qwen team. It is based on the Transformer architecture and incorporates technologies such as the SwiGLU activation function, attention QKV bias, group query attention, a mixture of sliding window attention, and full attention. Additionally, the Qwen team has improved the tokenizer to better adapt to multiple natural languages and code.

Qwen2 is the new generation of large language model series introduced by the Qwen team. It is based on the Transformer architecture and incorporates technologies such as the SwiGLU activation function, attention QKV bias, group query attention, a mixture of sliding window attention, and full attention. Additionally, the Qwen team has improved the tokenizer to better adapt to multiple natural languages and code.

Yi-1.5 is an upgraded version of Yi. It continues pre-training on Yi using a high-quality corpus of 500B tokens and is fine-tuned on 3M diverse samples.