DeepSeek-V2-0517 update boosts instruction-following by 14 points
JSON parsing accuracy hits 97% with regex support in latest DeepSeek model.
DeepSeek has rolled out an important update to its V2 model, dubbed DeepSeek-V2-0517, released on May 17, 2026. The update focuses on two core capabilities: instruction-following and structured output parsing. On the IFEval Benchmark, which measures how well a model follows complex prompts, the Prompt-Level accuracy surged from 63.9% to 77.6% — a 13.7 percentage point improvement. This means the model is now significantly better at handling multi-step instructions, conditional logic, and nuanced user requests without losing coherence.
For JSON parsing, a critical requirement for API integrations and data extraction, DeepSeek-V2-0517 improved raw accuracy from 78% to 85%. More notably, when combined with regular expressions, the accuracy reaches up to 97%. This enhancement reduces errors in structured data generation, making the model more robust for automated workflows. The update also strengthens the model's performance in complex coding tasks and natural language understanding, positioning it as a stronger competitor in the open-weight LLM space.
- IFEval Benchmark Prompt-Level accuracy improved from 63.9% to 77.6%, a 13.7 point gain.
- JSON parsing accuracy rose from 78% to 85%, and up to 97% with regular expressions.
- Update enhances reliability for complex coding and NLP tasks, reducing output errors.
Why It Matters
Better instruction-following and JSON parsing make DeepSeek more reliable for production code and automation workflows.