DeepSeek-V2.5

DeepSeek-V2.5 was released on September 5, 2024 as a merger of the DeepSeek-V2 chat line and DeepSeek-Coder-V2, creating a single model that handles both general conversation and programming. It kept the V2 architecture (236B total parameters, 21B active per token, multi-head latent attention) while substantially improving alignment: the release notes report ArenaHard win rate rising from 68.3% to 76.3% and MT-Bench from 8.84 to 9.02, alongside stronger code generation carried over from the coder lineage.

At launch it was widely covered as one of the strongest openly available models of its moment, and it served as the backbone of DeepSeek's hosted API (via the deepseek-chat and deepseek-coder endpoints) from September until the V3 transition in December 2024. The weights are published on Hugging Face under DeepSeek's custom model license, which permits commercial use but is not an OSI-approved open-source license.