Frontier Log

DeepSeek

DeepSeek-V4-Flash

Announced on deepseek.com

Benchmark results

BenchmarkScoreSource check
SWE-bench Verifiedmax reasoning effort79.0%
Terminal-Bench 2.0max reasoning effort56.9%
MMLU-Promax reasoning effort86.2%
LiveCodeBenchmax reasoning effort; problem window not stated91.6%
MCP Atlasmax reasoning effort; public set69.0%
Sources

Reported by DeepSeek

Related events

All models