AI Trend Notifier
EN
← wiki

$ cat wiki/models/deepseek-v4.md

DeepSeek V4

modelupdated 2026-08-14created 2026-07-20

Compared with

Spec

AttributeValue
DeveloperDeepSeek
Released2026-07-16 (GA; preview: 2026-04-24)
Announced2026-04-24
Context window1M tokens (384K max output)
PricingPeak: 2× off-peak API rates; off-peak rates not disclosed
LicenseMIT (open-weight)
AvailabilityDeepSeek API, Hugging Face (deepseek-ai/DeepSeek-V4-Pro)

Release Date

Preview: April 24, 2026. GA: mid-July 2026.

Benchmarks

  • DeepSeek-V4-Pro-Max: 80.6% SWE-bench Verified — highest open-weights score, tied with Gemini 3.1 Pro

Variants

VariantTotal ParamsActive ParamsUse Case
V4-Pro1.6T49BQuality-sensitive reasoning
V4-Flash284B13BFaster, lower-cost serving
V4-Pro's GA build now has its own page — DeepSeek V4-Pro-0813. On
2026-08-13 DeepSeek replaced the preview edition of V4-Pro in its API docs with build
0813, adding Responses-API and Codex support and reporting large agentic gains
against the preview; the independent Artificial Analysis index moved by one point
(source). Everything on this
page describing V4-Pro refers to the July GA / preview state.

V4-Flash now has its own page — DeepSeek V4-Flash. It was released officially on 2026-07-31 as build 0731, a post-training-only refresh of the same 284B/13B architecture, and on vendor-stated agent benchmarks it overtakes V4-Pro-Preview (Terminal Bench 2.1 82.7 vs 72.1) (source). The same announcement states the official V4-Pro release "will follow soon", which this page does not yet record as shipped.

Architecture

  • Mixture-of-Experts (MoE)
  • Hybrid Compressed Sparse Attention (CSA) + Heavily Compressed Attention (HCA)
  • 1M-token context at only 27% of single-token inference FLOPs (vs V3.2)
  • Only 10% of KV cache footprint of V3.2 at 1M tokens
  • Dual mode: Thinking (chain-of-thought) / Non-Thinking

Use Cases

  • Agentic coding (SWE-bench frontier open-weight)
  • Long-context reasoning
  • Cost-efficient frontier tasks

Deprecation Notice

  • deepseek-chat and deepseek-reasoner (legacy models) retired July 24, 2026 15:59 UTC

Compared To

ModelSWE-bench VerifiedOpen?
DeepSeek V4-Pro-Max80.6%Yes (MIT)
Claude Fable 580.3% (SWE-bench Pro)No
Gemini 3.1 Pro80.6%No
GLM-5.2unknownYes (MIT)

Sources

Referenced by

Sources