DeepSeek V4 Flash (0731): benchmark results
DeepSeek's July 31, 2026 official release of V4 Flash, an open-weight MoE model superseding the earlier V4 Flash preview with production-focused efficiency and agentic improvements. Provider: DeepSeek. Released 2026-07-31. Access: Open.
Aggregate Rating 1780 ± 6, rank #84 of 1635 rated models, from 423 benchmark results.
Strongest benchmark results
| Benchmark | Score | Metric | Percentile |
|---|---|---|---|
| BenCzechMark - Czech Language Understanding | 92.52 | Duel Win Score (%) | 100 |
| BenCzechMark - Czech Math Reasoning | 93.59 | Duel Win Score (%) | 100 |
| CyberGym | 93.17 | Level 1 successful PoC rate (%) (self-reported) | 100 |
| HWE-Bench (OpenHands SDK) | 76.5 | Resolved (%) | 100 |
| NVIDIA Nemotron 3.5 Super VL Card - AA-LCR v1.1 | 79.7 | Score (%) | 100 |
| NVIDIA Nemotron 3.5 Super VL Card - AA-Omniscience accuracy | 45 | Score (%) | 100 |
| NVIDIA Nemotron 3.5 Super VL Card - CritPt | 12.6 | Score (%) | 100 |
| NVIDIA Nemotron 3.5 Super VL Card - LiveCodeBench v6 | 94.6 | Score (%) | 100 |
| NVIDIA Nemotron 3.5 Super VL Card - MMLU-ProX | 85.5 | Score (%) | 100 |
| NVIDIA Nemotron 3.5 Super VL Card - SciCode, subtask | 49.3 | Score (%) | 100 |
| NVIDIA Nemotron 3.5 Super VL Card - tau2-bench airline | 83.5 | Score (%) | 100 |
| Nejumi 4 - jaster (2-shot) - JCoLA (out-of-domain) | 91 | Exact match (%) | 100 |
Interactive version: theaggregate.ai/model?slug=deepseek-v4-flash-0731 · How It Works · Data refreshed daily, snapshot 2026-11-05.