DeepSeek V2.5

DeepSeek-V2.5 is an open-weights model from DeepSeek that combines two earlier models: DeepSeek-V2-0628 (chat) and DeepSeek-Coder-V2-0724 (code). DeepSeek announced it on September 5, 2024. It keeps the chat model's conversational ability and the coder model's code skills, with better alignment to human preferences, writing and instruction following. The weights are on the DeepSeek-V2.5 Hugging Face card.

Status

At the time of the announcement, DeepSeek served V2.5 through the backward-compatible deepseek-chat and deepseek-coder API names. DeepSeek's current pricing page lists newer model names (deepseek-flash and deepseek-v4-pro) and does not mention V2.5. Treat V2.5 as a superseded open-weights release, and use the Hugging Face weights if you need it. See DeepSeek V4 Flash for a current DeepSeek API model in this catalog.

Key Features

• Merged chat and code model: one model replaces the separate V2-Chat and Coder-V2 lines.
• 236B parameters: the Hugging Face card lists 236B parameters in BF16.
• API features unchanged: function calling, FIM completion and JSON output mode, according to DeepSeek's announcement.
• Vendor-reported results: DeepSeek's card lists AlpacaEval 2.0 at 50.5, ArenaHard at 76.2, HumanEval (Python) at 89 and LiveCodeBench (01-09) at 41.8. These are DeepSeek's own evaluations. The announcement also says its overall safety score rose from 74.4% to 82.6%.
• License: code under MIT; model use is governed by a separate Model License that the card says supports commercial use.

Use Cases

• Research or fine-tuning on a 236B open-weights chat and code base.
• Reproducing results that were reported against V2.5.
• Self-hosted coding assistants where V2.5 weights are already deployed.

Limitation: running BF16 inference needs eight 80 GB GPUs per the card, and the card says it is not deployed by any Hugging Face Inference Provider.

Pricing

The weights are free to download under the license above. This page does not quote API prices because DeepSeek's current pricing page does not list V2.5.

Getting Started

1. Read the Hugging Face card and its Model License.
2. Load with Transformers using trust_remote_code=True, or serve with vLLM as the card describes (it requires a specific pull request, vllm-project/vllm#4650).
3. For hosted use, pick a current model from DeepSeek's API docs.

Frequently Asked Questions

When was DeepSeek-V2.5 released?

DeepSeek announced it on September 5, 2024.

Is DeepSeek-V2.5 open source?

The weights are on Hugging Face. Code is MIT-licensed and model use follows DeepSeek's Model License, which the card says supports commercial use.

What replaced DeepSeek-V2.5?

DeepSeek's current pricing page lists newer models such as deepseek-v4-pro and deepseek-flash. It does not name a formal successor to V2.5.

How big is V2.5?

236B parameters, according to the Hugging Face card.

Alternatives

• DeepSeek V4 Flash: a current DeepSeek model.
• DeepSeek V3: the next major DeepSeek chat release in this catalog.
• DeepSeek Coder V2.5: the code-focused sibling entry.

Conclusion

DeepSeek-V2.5 merged DeepSeek's chat and coder lines in September 2024. It is still downloadable, but DeepSeek's API has moved to newer models, so choose it only for self-hosting or reproducibility. Details are on the official announcement.