DeepSeek-V4-Flash-FP8: benchmark results

DeepSeek V4 Flash evaluated in FP8 using the sgl-project conversion served by SGLang. The download links to that FP8 artifact. Provider: DeepSeek. Released 2026-04-24. Access: Open.

Aggregate Rating 1698 ± 27, rank #358 of 10916 rated models, from 15 benchmark results.

Strongest benchmark results

BenchmarkScoreMetricPercentile
GSM-MC99.47Accuracy (%)97.8
RewardBench 2 Precise IF68.91Score (%)95.2
RewardBench 2 Focus93.64Score (%)93.5
RewardBench 2 Math88.93Score (%)92.1
RewardBench 2 Factuality76.68Accuracy (%)85.1
MATH-MC Level 599.53Accuracy (%)78.7
MATH-MC Level 499.33Accuracy (%)75
RewardBench 2 Safety87.56Score (%)73.6
JudgeBench Coding96.43Accuracy (%)73.5
MATH-MC Level 398.92Accuracy (%)65.4
JudgeBench Math89.29Accuracy (%)60.8
JudgeBench Knowledge77.6Accuracy (%)56.9

Interactive version: theaggregate.ai/model?slug=deepseek-v4-flash-fp8 · How It Works · Data refreshed daily, snapshot 2026-11-08.