.png)
November 20, 2024
DeepSeek, an emerging company in the field of artificial intelligence, recently announced the release of DeepSeek-V2.5. This new version combines the strengths of DeepSeek-V2-Chat and DeepSeek-Coder-V2-Instruct, integrating general language capabilities with specialized coding skills. The announcement was made via a tweet from @_akhaliq, pointing to the model's availability on Hugging Face.
DeepSeek-V2.5 aims to improve usability and performance compared to its predecessors. The model has been optimized in various areas, including writing style and adherence to instructions. The developers emphasize the improved alignment with human preferences. In benchmarks like AlpacaEval 2.0, ArenaHard, AlignBench, and MT-Bench, DeepSeek-V2.5 shows improvements over DeepSeek-V2-0628 and DeepSeek-Coder-V2-0724. Progress has also been made in the area of code generation, measured by HumanEval Python and LiveCodeBench.
Running DeepSeek-V2.5 locally in BF16 format requires powerful resources, specifically 80GB*8 GPUs. Inference can be performed directly with Huggingface Transformers. Correct configuration of parameters such as max_memory and device_map is crucial. Alternatively, the use of vLLM is recommended, incorporating a specific pull request for optimal compatibility.
DeepSeek-V2.5 offers advanced features like function calls, allowing the model to utilize external tools. This expands the AI's capabilities and enables more complex applications. JSON output enables the generation of structured data, which is particularly beneficial for integration with other systems. FIM (Fill In the Middle) completion allows the model to be provided with a prefix and optionally a suffix, with the AI generating the content in between.
The repository's code is licensed under the MIT License. Use of the DeepSeek-V2 Base/Chat models is subject to the Model License. The DeepSeek-V2 series, including Base and Chat, supports commercial use. For questions, users can create an issue or contact service@deepseek.com via email.
The development of DeepSeek-V2.5 fits into the dynamic landscape of large language models. The focus on economical training and efficient inference addresses important challenges in this field. The combination of general language capabilities with specialized coding skills positions DeepSeek-V2.5 as a versatile tool for various applications.
Bibliographie: https://www.deepseek.com/ https://huggingface.co/deepseek-ai/DeepSeek-V2.5 https://huggingface.co/spaces/akhaliq/anychat/discussions/1 https://github.com/deepseek-ai/DeepSeek-V2 https://deepinfra.com/deepseek-ai/DeepSeek-V2.5 https://x.com/_akhaliq?lang=de https://twitter.com/deepseek_ai https://www.reddit.com/r/LocalLLaMA/comments/1fclav6/all_of_this_drama_has_diverted_our_attention_from/