DeepSeek · open · 2025-12
DeepSeek-V3.2
#12 index 78.8
the takeThe workhorse that made cheap long context normal with sparse attention. Still fine, but the first-party API is being retired in weeks and V4 is simply better. A baseline, not a buy.
SWE-bench Verified
70.0
GPQA Diamond
82.4
MMLU-Pro
85.0
AIME (math)
84.1
- Context
- 164K
- Input
- $0.28/M
- Output
- $0.42/M
- Speed
- —
- Modality
- text
- Efficient long-context via sparse attention
- Very low cost, open weights (MIT)
- Strong general knowledge and tool use
- Superseded by the V4 family
- 163K context vs V4 1M
- First-party API deprecating 2026-07-24