Select Language:
DeepSeek AI Unveils Upgraded 2.5 Model: Merging Coder and Chat for Enhanced Performance
On September 6, 2023, DeepSeek AI announced the launch of its latest model upgrade, DeepSeek V2.5, which merges the capabilities of the previously separate models: DeepSeek Coder V2 and DeepSeek V2 Chat. This new model promises to align more closely with human preferences and optimize writing tasks and instruction adherence.
In a recent update, DeepSeek released its API support documentation, confirming the integration of the two earlier models into a single, enhanced version. API users will be able to access the new DeepSeek V2.5 using either “deepseek-coder” or “deepseek-chat,” ensuring backward compatibility.
The updated model significantly surpasses the original versions in both general capabilities and code proficiency. Key improvements include:
- ArenaHard win rate increasing from 68.3% to 76.3%.
- AlpacaEval 2.0 LC win rate rising from 46.61% to 50.52%.
- MT-Bench score improving from 8.84 to 9.02.
- AlignBench score advancing from 7.88 to 8.04.
Furthermore, DeepSeek V2.5 builds on the original Coder model by further enhancing code generation capabilities, specifically optimizing for common programming scenarios. The new model achieved impressive results on standard testing sets, including a HumanEval score of 89% and a LiveCodeBench score of 41% from January to September.
DeepSeek AI, founded in 2023 and based in Hangzhou, is focused on researching advanced general artificial intelligence models and technologies. Since its inception, the company has rapidly developed and open-sourced multiple large models with billions of parameters, including the DeepSeek-LLM general language model and the DeepSeek-Coder model. In January 2024, they were the first in the country to open-source a Mixture of Experts (MoE) model, demonstrating exceptional performance compared to peers on public evaluation benchmarks.
This innovative upgrade underscores DeepSeek’s commitment to pushing the boundaries of AI technology while improving user experience and efficiency in applications ranging from coding to conversational AI.
ChatGPT
Perplexity
Gemini AI
Grok AI
