← Back to directory
DeepSeek
DeepSeek-V3.2
Released Sep 29, 2025 · knowledge cutoff 2025-07
- Status
- Superseded
- Location
- China
- Type
- LLM
- Modality
- Text
- Context window
- 128K
- Max output
- 66K
- Speed
- —
- Price ($/MTok in / out)
- $0.28 / $0.42
- Cost per task
- —
- Open weights
- Yes
Benchmarks7 of 10 reported
MMLU-Proretired
85%
GPQA Diamond
79.9%
SWE-bench Verified
67.8%
SWE-bench Pro
15.56%
AIME (latest)retired
89.3%
Humanity's Last Exam
19.8%
LMArena Elo
1420
Terminal-Bench 2.1
—
GDPval-AA v2
—
ARC-AGI-2
—
Strengths & weaknesses
Strengths
- Sparse attention keeps long-context work cheap instead of scaling quadratically
- Cheap enough to leave running on background reasoning work
- Open weights under an unrestrictive licence
Weaknesses
- An experimental release — behavior shifted noticeably between checkpoints
- Prose and instruction-following still trail its reasoning
Listed as V3.2-Exp at release. SWE-bench Pro 15.56% is Scale's public-set leaderboard figure for "deepseek-v3p2" — third-party, not DeepSeek-reported.
For developers
API model strings
Not researched
Licence
MIT License · Permissive · commercial use permitted
Retirement
No retirement announced
Lineage
No recorded predecessor