news

한어Русский языкEnglishFrançaisIndonesianSanskrit日本語DeutschPortuguêsΕλληνικάespañolItalianoSuomalainenLatina

original title: deepseek ai model upgrade released version 2.5: merge coder and chat, align human preferences, optimize writing tasks and command following, etc.

it home reported on september 6 that the two models, deepseek coder v2 and deepseek v2 chat, have been merged and upgraded, and the upgraded new model is deepseek v2.5.

image source: screenshot from linux.do forum user

deepseek officially updated its api support documentation yesterday (september 5), announcing the merger of the deepseek coder v2 and deepseek v2 chat models, and the launch of a new deepseek v2.5 model.

officials said that for forward compatibility, api users can access the new model through deepseek-coder or deepseek-chat.

the new model significantly surpasses the two old models in terms of general capabilities and coding capabilities.

the new model better aligns with human preferences and is optimized in many aspects, such as writing tasks and command following:

arenahard winrate increased from 68.3% to 76.3%

alpacaeval 2.0 lc winrate increased from 46.61% to 50.52%

mt-bench score increased from 8.84 to 9.02

alignbench score increased from 7.88 to 8.04

the new model further improves the code generation capability based on the original coder model, optimizes common programming application scenarios, and achieves the following results on the standard test set:

HumanEval: 89%

livecodebench (january-september): 41%

it home note:the deepseek ai model was launched by hangzhou deepseek artificial intelligence, which was founded in 2023.

the official introduction is as follows:

focusing on researching the world's leading general artificial intelligence underlying models and technologies, and challenging cutting-edge problems in artificial intelligence. based on self-developed training frameworks, self-built intelligent computing clusters, and wanka computing power, the deepseek team has released and open-sourced multiple large models with tens of billions of parameters in just half a year, such as the deepseek-llm general language model and the deepseek-coder code model. in january 2024, it took the lead in open-sourcing the first moe large model (deepseek-moe) in china. the generalization effects of various models outside the public evaluation list and real samples have all surpassed the outstanding performance of models of the same level.