Atlas

Models

← All models

DeepSeek-V4-Flash

DeepSeek · 2026-04-24 · 292B parameters

DeepSeek-V4-Flash is the smaller, efficiency-focused open-weight MoE model in DeepSeek's V4 preview release, marketed at 284B total and 13B activated parameters. Released April 24, 2026, it supports a 1M-token context and Non-Think, Think High, and Think Max reasoning modes; its 43-layer attention stack begins with two sliding-window layers, then interleaves Compressed Sparse Attention (CSA) and Heavily Compressed Attention (HCA). Its weights are released under the MIT license.

Benchmark scores

BenchmarkScore
A-Fantasia73.3
A-Fantasia Backwards Spelling87.0
A-Fantasia Chess53.0
A-Fantasia Cube Rotation80.0
AA-Briefcase Elo833
AA-LCR63.0
AA-Omniscience Index-22.3
AIME 202695.8
ALE-Bench1380
Antidote: Everyday Edition968
Apex (MathArena)27.1
Apex Shortlist89.4
Arena Code WebDev — Brand & Marketing1597
Arena Code WebDev — Consumer Product1553
Arena Code WebDev — Data & Analytics1527
Arena Code WebDev — Gaming1601
Arena Code WebDev — HTML1527
Arena Code WebDev — React1590
Arena Code WebDev — Reference-Based Design1573
Arena Code WebDev — Simulations1581
Arena Text — Business, Management & Financial Ops (No Style Control)1425
Arena Text — Chinese (No Style Control)1468
Arena Text — Creative Writing (No Style Control)1402
Arena Text — English (No Style Control)1441
Arena Text — Entertainment, Sports & Media (No Style Control)1401
Arena Text — Exclude Ties (No Style Control)1424
Arena Text — Exclude Ties (Style Controlled)1431
Arena Text — Expert (No Style Control)1444
Arena Text — French (No Style Control)1434
Arena Text — French (Style Controlled)1444
Loading Atlas data…