DeepSeek

DeepSeek-V4-Flash-0731

Released Jul 31, 2026 · knowledge cutoff unpublished

Status: Deprecated
Location: China
Type: LLM
Modality: Text
Context window: 1M
Max output: 393K
Speed (max effort): 123 tok/s · 1.34s to first answer token
Price ($/MTok in / out): $0.22 / $0.66
Cost per task (max effort): $0.027
Open weights: Yes

Benchmarks4 of 10 reported

GPQA Diamond

90.8%

Terminal-Bench 2.1

82.7%

Humanity's Last Exam

36.8%

ARC-AGI-2

61.4%

MMLU-Proretired

—

SWE-bench Verified

—

SWE-bench Pro

—

AIME (latest)retired

—

LMArena Elo

—

GDPval-AA v2

—

Strengths & weaknesses

Strengths

Same weights and architecture as the April preview — every gain comes from a rebuilt post-training recipe, not new pre-training
Clears DeepSeek's own larger V4-Pro preview on every agentic benchmark the lab published
Three reasoning-effort levels and a speculative-decoding module carried inside the one checkpoint
MIT-licensed and ungated, so on-premise commercial deployment needs no negotiation

Weaknesses

Thinks at length before acting on max effort — long agent runs burn far more tokens and wall-clock than the per-call latency suggests
Fabricates heavily on knowledge probes, and factual coverage has not moved with the agentic gains
Ships without a Jinja chat template — messages go through DeepSeek's own Python encoding scripts instead

The production release of DeepSeek-V4-Flash, superseding April's preview: same 284B/13B MoE and DSpark speculative-decoding module, with all gains from re-post-training. Terminal-Bench 2.1 82.7 is DeepSeek's own figure at max reasoning effort in its minimal-mode harness; Artificial Analysis measures 78.7 on the same benchmark. GPQA Diamond 90.8 and HLE 36.8% (no tools) are Artificial Analysis' measurements — DeepSeek published neither. Also reported by DeepSeek: NL2Repo 54.2, Cybergym 76.7, DeepSWE 54.4, Toolathlon-Verified 70.3, Agents' Last Exam 25.2. MMLU-Pro, SWE-bench Verified, AIME and LMArena are unpublished; ARC-AGI-2 comes from ARC Prize (below). Artificial Analysis scores it 50 on its Intelligence Index, 10 points above the April preview’s 40. Max output is the 384K DeepSeek recommends for high and max effort; the OpenRouter endpoint caps completions at 65,536. Pricing updated 2026-09-03: DeepSeek moved the V4 line to peak/off-peak billing on 2026-08-16 16:00 UTC (announced 2026-08-14) — the flat $0.14/$0.28 became $0.22/$0.66 off-peak and $0.44/$1.32 during peak hours (01:00-04:00 and 06:00-10:00 UTC); the off-peak rate is recorded here as the base tier, and costPerTask ($0.0271) predates the change. Retired 2026-09-10: DeepSeek's change log states "the previous-generation models V4 Flash and V4 Flash Vision Exp have been retired" with the launch of DeepSeek-V4.1-Flash; the `deepseek-v4-flash` id is still accepted but temporarily routed to V4.1-Flash and billed at Flash rates, so calls to it no longer reach this model. Weights remain on Hugging Face. ARC-AGI-2 61.4% is ARC Prize-verified at max effort ($0.04/task; high 56.0%, low 46.0%), published 2026-07-31 at arcprize.org/results/deepseek-v4-flash-0731.

For developers

API model strings

deepseekdeepseek-v4-flash

Licence

MIT License · Permissive · commercial use permitted

Retirement

Sep 10, 2026

Lineage

No recorded predecessor