DeepSeek-V4-Flash-Vision-Exp: benchmark results
DeepSeek's open-weight DeepSeek-V4-Flash-Vision-Exp (August 2026), the DeepSeek-V4 family's first experimental multimodal model, built on the V4 Flash architecture with added vision capabilities. Provider: DeepSeek. Released 2026-08-21. Access: Open.
Aggregate Rating 1785 ± 15, rank #76 of 1656 rated models, from 169 benchmark results.
Strongest benchmark results
| Benchmark | Score | Metric | Percentile |
|---|---|---|---|
| LiveBench Python | 75 | Score | 97.1 |
| Nejumi 4 - jaster (2-shot) - JCoLA (out-of-domain) | 89 | Exact match (%) | 94.8 |
| Nejumi 4 - jaster (0-shot) - CommonsenseMoralJA | 97 | Exact match (%) | 93.7 |
| Nejumi 4 - ARC-AGI-1 | 98 | Accuracy (%) | 93.4 |
| Korean CSAT 2026 (Easy Mode) - Korean | 100 | Points (out of 100) | 92.6 |
| LiveBench TypeScript | 56.67 | Score | 90.7 |
| Korean CSAT 2026 (Easy Mode) - English | 100 | Points (out of 100) | 90.5 |
| Nejumi 4 - BFCL - Relevance Detection | 88.89 | Accuracy (%) | 89.5 |
| LLM Stats Score | 44.03 | LLM Stats Score (conservative rating) | 87.5 |
| Nejumi 4 - HLE (Japanese) - Calibration Error | 53 | Calibration error (%) | 87.4 |
| Nejumi 4 - jaster (0-shot) - NIILC | 57.72 | Character F1 (x100) | 87.4 |
| Nejumi 4 - HLE (Japanese) - Accuracy | 38.14 | Accuracy (%) | 86.7 |
Interactive version: theaggregate.ai/model?slug=deepseek-v4-flash-vision-exp · How It Works · Data refreshed daily, snapshot 2026-10-20.