Atlas

Models

← All models

DeepSeek-V4-Flash-0731

DeepSeek · 2026-07-31

DeepSeek-V4-Flash-0731 is the official July 31, 2026 release of DeepSeek-V4-Flash, superseding the April preview after a retrained post-training pipeline focused on coding, agents, reasoning, and tool use. It keeps the preview's marketed 284B-total/13B-activated MoE design with a 1M-token context and Non-Think, Think High, and Think Max modes; its published checkpoint config matches the preview's architecture (43 layers, 4096 hidden size, 64 attention heads, 129280 vocabulary). Its weights are released under the MIT license.

Benchmark scores

BenchmarkScore
AA-LCR65.7
AA-Omniscience Index-15.7
Artificial Analysis Agentic Index45.7
Artificial Analysis Coding Index69.1
Artificial Analysis Intelligence Index49.9
Artificial Analysis Omniscience Accuracy37.2
Artificial Analysis Omniscience Hallucination Rate84.4
Capability158.6
CritPt16.6
DeepsecBench16.5
GDPval-AA v21559
GDPval-AA v2 (normalized)52.9
GPQA Diamond90.8
Humanity's Last Exam (Text-Only)36.8
Ramp SWE-Bench79.5
SciCode (Artificial Analysis)49.9
Terminal-Bench 2.178.7
𝜏³-Banking31.1
Loading Atlas data…