DeepSeek merges chat and coder lines into DeepSeek-V2.5

DeepSeek combined its separate general-purpose chat and coder model lines into a single checkpoint, DeepSeek-V2.5, merging DeepSeek-V2-0628 and DeepSeek-Coder-V2-0724 into one model that retained both the earlier chat model’s conversational ability and the coder model’s code-processing strength.

The company reported broad gains from the merge: its ArenaHard win rate, a measure of how often the model’s responses were preferred head-to-head against a reference model, rose from 68.3% to 76.3%, alongside smaller improvements on AlpacaEval 2.0’s length-controlled win rate (46.61% to 50.52%) and MT-Bench (8.84 to 9.02). DeepSeek said the update improved writing quality and instruction-following in addition to coding ability, rather than trading one capability off against the other as a naive merge might. For continuity, the merged model remained accessible through both the existing deepseek-chat and deepseek-coder API endpoints, so integrations built against either did not need to change.

The release was a comparatively minor step in DeepSeek’s public output through 2024, but it reflected a broader pattern in the company’s development: consolidating specialised checkpoints into fewer, more general models as capability improved, rather than maintaining an expanding set of task-specific variants — the same direction other frontier labs converged on as “one model that does everything well” became more achievable than it had been a year earlier.