DeepSeek-V2.5

What it is

DeepSeek-V2.5 is a model DeepSeek released on 5 September 2024^[1]. Its main change is not scale but consolidation: the chat model DeepSeek-V2-0628 and the code model DeepSeek-Coder-V2-0724 were merged into a single model that keeps the chat model's conversational ability and the coder model's handling of code^[1].

Where it came from, and how it is used

Before the merge, choosing a model meant choosing between conversation and code. Afterwards one model covers both, and the way callers reach it does not change: it is served through both the deepseek-chat and deepseek-coder API model names, with Function Calling, FIM completion and JSON output unchanged^[1]. Function calling lets a model invoke external tools, FIM (fill-in-the-middle) completion lets it fill a gap in a piece of code from the surrounding context, and JSON output makes replies machine-readable. Existing integrations keep working; what changes is the model answering behind them.

What DeepSeek measured

DeepSeek reported results in two areas.

Safety. On its internal tests, the overall safety score rose from 74.4% with V2-0628 to 82.6%, and the safety spillover rate fell from 11.3% to 4.6%^[1]. Spillover here refers to safety behaviour bleeding into ordinary requests, so a lower figure means the model got better at turning down harmful requests without dragging normal ones down with them.

Code and completion. The model improved 5.1% on DeepSeek's internal fill-in-the-middle test set DS-FIM-Eval, and also improved on HumanEval Python and LiveCodeBench, but trailed Coder-V2-0724 slightly on HumanEval Multilingual and Aider^[1]. HumanEval is a set of programming problems where the model writes functions from docstrings; LiveCodeBench and Aider test code work in more interactive settings; SWE-verified tests fixes to real software repositories. The benchmarks named in the release are HumanEval Python, LiveCodeBench, HumanEval Multilingual, Aider, SWE-verified, DS-FIM-Eval and DS-Arena-Code^[1].

The release also says V2.5 brought significant improvements in writing and instruction-following and aligned better with human preferences^[1].

Availability

Weights were open-sourced on Hugging Face, and the model went live on DeepSeek's web chat and API^[1]. DeepSeek advised adjusting the system prompt and temperature settings for the best results if performance drops in certain cases^[1].

Where it sits in the line

V2.5 is one point in the V2 series. DeepSeek said the V2 series had delivered five updates since May 2024^[2]. DeepSeek-V2.5-1210, released on 10 December 2024, improved math, coding, writing and roleplay and closed out the V2 series^[2]; real-time Internet Search launched on chat.deepseek.com alongside it^[2]. DeepSeek-V3 succeeded it on 26 December 2024^[3].

The significance of V2.5 is mostly a matter of product shape. By folding chat and code into one model while keeping both API names compatible, it removed a choice from users and left DeepSeek maintaining a single main line of iteration rather than two.