DeepSeek

DeepSeek-V4

Released Apr 24, 2026 · knowledge cutoff unpublished

Status: Superseded
Location: China
Type: LLM
Modality: Text
Context window: 1M
Max output: 384K
Speed: —
Price ($/MTok in / out): $0.43 / $0.87
Cost per task (max effort): $0.045
Open weights: Yes

Benchmarks6 of 10 reported

MMLU-Proretired

87.5%

GPQA Diamond

90.1%

SWE-bench Verified

80.6%

SWE-bench Pro

55.4%

Terminal-Bench 2.1

72.1%

Humanity's Last Exam

48.2%

AIME (latest)retired

—

LMArena Elo

—

GDPval-AA v2

—

ARC-AGI-2

—

Strengths & weaknesses

Strengths

Cheap enough to change what you are willing to spend tokens on at all
Whole codebases fit in context, so cross-file work needs no chunking
The small variant is fast and good enough for most work; the large one is there when it isn't

Weaknesses

Reads as benchmark-tuned — stronger on tests than in day-to-day use
Serving capacity is chip-constrained, so the top variant is frequently unavailable
No dedicated reasoning line to escalate to

V4-Pro is a 1.6T/49B-active MoE; V4-Flash costs $0.14/$0.28. MMLU-Pro 87.5 / GPQA 90.1 are DeepSeek-reported for V4-Pro in Max reasoning mode; HLE 48.2 is third-party (LLM-Stats), same mode. Terminal-Bench 2.1 72.1 is DeepSeek's own figure for the April V4-Pro preview, published in its V4-Pro-0813 comparison table (harness not stated; the 0813 production build scores 87.9). SWE-bench Pro 55.4 ("SWE Pro") is DeepSeek's own model-card figure for V4-Pro in Max reasoning mode (huggingface.co/deepseek-ai/DeepSeek-V4-Pro; High 54.4, Non-Think 52.1).

For developers

API model strings

Not researched

Licence

MIT License · Permissive · commercial use permitted

Retirement

No retirement announced

Lineage

No recorded predecessor