DeepSeek merges its chat and coder models into DeepSeek-V2.5

Background

By mid-2024 DeepSeek had two lines of models in parallel. DeepSeek-V2 was a general chat model, and DeepSeek-Coder-V2 was a code model. Users who wanted both conversation and programming help had to choose between them, and developers had to manage two sets of model names in their applications.

Merging the two lines was a way to offer one model that could handle both kinds of work.

What happened

On 5 September 2024 DeepSeek released DeepSeek-V2.5 ^[1]. It merged DeepSeek-V2-0628, the chat model, and DeepSeek-Coder-V2-0724, the code model, into one model. The merged model kept the chat model's conversational ability and the coder model's code processing ^[1].

Weights were open-sourced on Hugging Face, and the model went live on DeepSeek's web chat and its API at the same time ^[1][2]. An open-weights release lets anyone download and run the model, while the web chat and API let people use it without running it themselves.

Key details

The model was served through both the deepseek-chat and deepseek-coder API names. Function Calling, fill-in-the-middle completion and JSON output were left unchanged ^[1]. Developers who had built on either earlier name could keep their existing integrations.

DeepSeek said the merged model's overall safety score on its internal tests rose from 74.4% for V2-0628 to 82.6%, and its safety spillover rate fell from 11.3% to 4.6% ^[1].

Impact

The merge reduced the number of separate models a user or developer had to manage, and it made a single open-weights model available for both conversation and code. For developers this means one model to test, monitor and route requests to instead of two, and for users the same model now covers both writing and programming, so they need not learn which product fits which task. Readers following the company's model history can treat this event as the point at which its chat and code work came together in one release.