Model

DeepSeek-V4-Pro (Non-Think)

OPEN

DeepSeek · DeepSeek-V4 family · released Apr 24, 2026

Fast/low-latency mode of the DeepSeek-V4-Pro preview checkpoint, the default reasoning tier a user gets without selecting anything. (fast/low-latency mode; the default tier a user gets without selecting anything).

ReasoningCodingVisionFunction callingTool useAgentic

1670.4

Elo · rank #91

Parameters

1600B

Active params

49B (MoE)

Context

1M tokens

Architecture

Mixture-of-Experts (1.6T total / 49B active), hybrid Compressed Sparse Attention + Heavily Compressed Attention, Manifold-Constrained Hyper-Connections, Muon optimizer

License

MIT

Languages

—

API price (in/out)

$0.435 / $0.87

Modalities

text

Benchmark results

Bar shows position within the tracked field; marker = field best

GPQA DiamondReasoning72.9%#113

best: GPT-5.6 · 94.6%

Humanity's Last ExamReasoning7.7%#101

best: Claude Sonnet 5 · 57.4%

IFBenchReasoning46.0%#39

best: MiniMax M3 · 83.0%

LiveCodeBenchCoding56.8%#105

best: DeepSeek-V4-Pro (Think Max) · 93.5%

MMLU-ProKnowledge82.9%#41

best: Claude Fable 5 · 91.5%

SWE-bench ProCoding52.1%#34

best: Claude Fable 5 · 80.0%

SWE-bench VerifiedCoding73.6%#47

best: Claude Fable 5 · 95.0%

Terminal-Bench 2.0Coding59.1%#35

best: GPT-5.6 · 88.8%

Run it locally

VRAM @ Q4

—

VRAM @ FP16

—

Fits on (Q4)

Multi-node cluster required

Throughput data unavailable.

Quantizations

—

DeepSeek-V4 family

Elo progression across releases

DeepSeek-V4-Flash (Non-Think)Apr 20261623.2 DeepSeek-V4-Flash (Think High)Apr 20262239.5 DeepSeek-V4-Flash (Think Max)Apr 20262330.6 DeepSeek-V4-Pro (Non-Think)Apr 20261670.4 DeepSeek-V4-Pro (Think High)Apr 20262395.1 DeepSeek-V4-Pro (Think Max)Apr 20262517.1

Resources

HFdeepseek-ai/DeepSeek-V4-Pro DOCShttps://api-docs.deepseek.com/news/news260424/APIAvailable via DeepSeek API

API price $0.435/$0.87 · each benchmark row carries its own source badge (see methodology)