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 and 1M-token context, supports low, high, and max reasoning effort, and keeps the preview's backbone configuration (43 layers, 4096 hidden size, 64 attention heads, 129280 vocabulary) while adding a DSpark speculative-decoding module. Its weights are released under the MIT license.

Benchmark scores

BenchmarkScore
A-Fantasia (Valid Attempts with Retries)80.7
A-Fantasia Chess (Valid Attempts with Retries)70.0
A-Fantasia Cube (Valid Attempts with Retries)78.0
A-Fantasia Spelling (Valid Attempts with Retries)94.0
AA-Briefcase Elo1253
AA-Briefcase v1.1 Elo1256
AA-LCR65.7
AA-LCR v1.179.7
AA-Omniscience Index-14.3
ALE-Bench1686
ARC-AGI-1 Semi-Private89.0
ARC-AGI-261.4
Artificial Analysis Agentic Index45.7
Artificial Analysis Coding Agent Index v1.538.7
Artificial Analysis Coding Index69.1
Artificial Analysis Intelligence Index49.9
Artificial Analysis Intelligence Index v4.334.5
Artificial Analysis Intelligence Index v4.3.234.3
Artificial Analysis Omniscience Accuracy40.4
Artificial Analysis Omniscience Hallucination Rate91.7
Artificial Analysis Openness Index44.4
Artificial Analysis Progress Index v1309
AutomationBench-AA 1.0.654.0
BioMysteryBench — Hard (Vals)23.5
BioMysteryBench — Human-solvable (Vals)74.0
BioMysteryBench (Vals Terminus 2)64.4
BullshitBench v2 Leaderboard20.0
CadQueryEval (Corrected Geometry)52.0
CadQueryEval Corrected Geometry Bounding Box60.0
CadQueryEval Corrected Geometry Chamfer60.0
Loading Atlas data…