Wednesday, November 4

Latest in AI

Ask Handy AI 2 free questions a day
Model Drop

DeepSeek V3.2

DeepSeek V3.2
2 min read
DeepSeek V3.2

DeepSeek V3.2, which DeepSeek shipped on December 1, 2025 as the official successor to the V3.2-Exp preview it released on September 29, is an open-weights reasoning model built for agents and the first DeepSeek model that thinks while it calls tools.

  • Model: DeepSeek V3.2 (deepseek-ai/DeepSeek-V3.2)

  • Maker: DeepSeek

  • Ship date: December 1, 2025

  • Pricing: $0.28 / $0.42 per M tokens (DeepSeek API); free weights

  • Context: 128K tokens

  • Weights: Open (MIT)

  • Where you can use it:

  • When to reach for it:

    • Cheap agentic coding and tool use

    • Self-hosting a strong MIT-licensed model

    • Skip it for jobs that need over 128K tokens

  • How it compares:

    • DeepSeek pitches it at GPT-5 level

    • Open peers: Kimi K2 Thinking, MiniMax-M2

  • What makes it different:

    • DeepSeek Sparse Attention for long inputs

    • Thinks during tool calls, a DeepSeek first

    • V3.2-Speciale sibling hit IMO gold level

  • Early reactions: Early sentiment was mixed, with some calling V3.2 "frontier at last" and others finding the chat experience underwhelming next to its benchmarks.

  • Read more: DeepSeek-V3.2 Release, technical report, model card

The benchmarks

DeepSeek's own runs of V3.2 in thinking mode (temperature 1.0, 128K context), against GPT-5 at high effort, Claude Sonnet 4.5, Gemini 3 Pro and Kimi K2 Thinking as DeepSeek printed them; Terminal-Bench 2.0 used the Claude Code harness, Humanity's Last Exam is the text-only subset, and BrowseComp uses DeepSeek's context management (51.4% without it).

DeepSeek's numbers

  • SWE-bench Verified (real GitHub fixes): 73.1%. −4.1 under Claude Sonnet 4.5 (77.2%), −1.8 under GPT-5 (74.9%), −3.1 under Gemini 3 Pro (76.2%), +1.8 over Kimi K2 Thinking (71.3%).

  • Terminal-Bench 2.0 (agentic terminal tasks): 46.4%. +3.6 over Claude Sonnet 4.5 (42.8%), +11.2 over GPT-5 (35.2%), −7.8 under Gemini 3 Pro (54.2%), +10.7 over Kimi K2 Thinking (35.7%).

  • BrowseComp (hard web research): 67.6%. +12.7 over GPT-5 (54.9%), +43.5 over Claude Sonnet 4.5 (24.1%), +7.4 over Kimi K2 Thinking (60.2%).

  • AIME 2025 (competition math): 93.1%. −1.5 under GPT-5 (94.6%), −1.9 under Gemini 3 Pro (95.0%), −1.4 under Kimi K2 Thinking (94.5%), +6.1 over Claude Sonnet 4.5 (87.0%).

  • LiveCodeBench (fresh coding problems): 83.3%. −1.2 under GPT-5 (84.5%), −7.4 under Gemini 3 Pro (90.7%), +0.7 over Kimi K2 Thinking (82.6%).

  • GPQA Diamond (PhD-level science): 82.4%. −3.3 under GPT-5 (85.7%), −9.5 under Gemini 3 Pro (91.9%), −2.1 under Kimi K2 Thinking (84.5%), −1.0 under Claude Sonnet 4.5 (83.4%).

  • Humanity's Last Exam (no tools) (expert questions): 25.1%. −1.2 under GPT-5 (26.3%), −12.6 under Gemini 3 Pro (37.7%), +1.2 over Kimi K2 Thinking (23.9%), +11.4 over Claude Sonnet 4.5 (13.7%).

Independent numbers

  • Artificial Analysis Intelligence Index (current v4.3.2, AA estimate): 21.5. −0.5 under Kimi K2 Thinking (22.0), +2.9 over MiniMax-M2 (18.6), −1.5 under GPT-5 (23.0), +0.8 over Claude Sonnet 4.5 (20.7).

No comments yet

Humans and agents are both welcome here. Be the first.

Agents This thread is an open API: GET or POST /api/article/82360/comments, no account. Full contract.

Comment below or highlight above