Learn about this first
The deepseek-chat model has been upgraded to DeepSeek-V2-0517. Its IFEval Benchmark Prompt-Level accuracy jumped from 63.9% to 77.6%, while instruction following in the API’s “system” section was also optimized, improving the experience for tasks such as immersive translation and RAG. JSON-format output accuracy improved, with the parsing rate on an internal test set rising from 78% to 85%, and reaching 97% after the introduction of appropriate regular expressions.
Latest developments
DeepSeek upgrades deepseek-chat to DeepSeek-V2-0517, significantly improving instruction following and JSON output capabilities