Skip to content
AI Radar
DE
Curated AI Intelligence Est. 2026
Status: Preview

DeepSeek-V3.2-Exp

Provider: DeepSeek Family: DeepSeek-V3 Preview
Official docs
analytics

Capabilities

Status Preview DeepSeek
Release Sep 2025

Experimentelle Zwischenversion mit Sparse Attention.

Recommended for

  • arrow_right Evaluierung, Prototyping, frühe Tests

Use Cases

  • arrow_right Chat, Zusammenfassung, Umformulierung, Wissensarbeit

Technical specs

  • Model ID
    deepseek-v3-2-exp
  • Family DeepSeek-V3
  • Context window —
  • Release 9/29/2025
terminal Official docs

Benchmarks

Arena Performance · lmarena.ai

Chat #3 / 87
1 Claude Opus 4.6 1,476
4 Claude Sonnet 4.6 1,389
6 Kimi K2.5 1,355
2 Grok 4 Fast 1,338
7 GLM-5 1,318
3 DeepSeek-V3.2-Exp 1,300
8 Qwen3.5-397B-A17B 1,295
Coding #6 / 135
1 Claude Opus 4.6 2,002
2 Gemini 3.1 Pro Preview 1,861
7 Kimi K2.5 1,780
3 GPT-5.4 1,679
8 GLM-5 1,650
9 Qwen3.5-397B-A17B 1,590
10 MiniMax M2.5 1,440
6 DeepSeek-V3.2-Exp 1,210
Games #6 / 94
1 Claude Opus 4.6 1,550
2 Claude Sonnet 4.6 1,401
3 GPT-5.4 1,270
4 Gemini 3.1 Pro Preview 1,235
5 Grok 4 Fast 1,190
6 DeepSeek-V3.2-Exp 1,155
Math #5 / 75
1 Claude Opus 4.6 1,795
2 GPT-5.4 1,740
3 Gemini 3.1 Pro Preview 1,695
4 Qwen3.5-397B-A17B 1,660
5 DeepSeek-V3.2-Exp 1,620
6 Kimi K2.5 1,585
7 Claude Sonnet 4.6 1,550
Reasoning #7 / 80
1 Claude Opus 4.6 1,820
2 GPT-5.4 1,755
3 Gemini 3.1 Pro Preview 1,710
4 Kimi K2-Thinking-0905 1,680
5 Claude Sonnet 4.6 1,640
6 Qwen3.5-397B-A17B 1,590
7 DeepSeek-V3.2-Exp 1,555

HumanEval

Overall 69.2% #5/66 open_in_new

SWE-Bench Verified

Overall 75.1% #6/77 open_in_new

Terminal-Bench

Overall 51.7% #6/25 open_in_new

Timeline

History

4/1/2026 · 1d ago

First captured in AI Radar

DeepSeek

9/29/2025 · Sep 2025

Model release

DeepSeek-V3.2-Exp

open_in_new Source