Home › AI Models › DeepSeek-V4

DeepSeek-V4

AI Model Language Current

The efficiency school scales up: 1.6T parameters, MIT license, 1M context.

You're looking at an AI model record: who built it, what it can do, and the hardware and data behind it.

DeepSeek-V4 went live in preview on April 24, 2026 in Pro and Flash variants with API access and open weights under MIT. V4 Pro is a 1.6-trillion-parameter mixture-of-experts activating 49B per token, pairing MoE with sparse attention for a 1M-token context — and posting 80.6% on SWE-bench Verified (as reported in launch coverage), frontier-class coding from an open model. It extends the V3/R1 lineage's argument that engineering efficiency, not raw compute access, sets the pace under export controls.

Specifications & key facts

Each figure is labeled by basis: vendor spec, vendor claim, reported, estimate, or official. "Not publicly disclosed" means exactly that — we never invent values.

Overview

Architecture MoE + sparse attention; 1.6T total / 49B activeReported

Coding 80.6% SWE-bench VerifiedReported

How this connects to the rest of the stack

This record's links at a glance — click a group to jump to it.

Developed by

DeepSeek

Benchmarked on

SWE-bench Verified

Based on

DeepSeek-V3

Successor to

DeepSeek-V3

Competes with

GPT-5.6 (and the 2026 GPT-5.x line) Qwen 3.6 / 3.7-Max / 3.8 (2026 line)

Export Controls & the Chip War

Sources

Open-source LLM landscape 2026 (V4 details) ↗ · DeepSeek ↗

Record reviewed as of 2026-08-31 — see Methodology.