V3.2
DeepSeek · released Dec 1, 2025
DeepSeek-V3.2 is a large language model designed to harmonize high computational efficiency with strong reasoning and agentic tool-use performance. It introduces DeepSeek Sparse Attention (DSA), a fine-grained sparse attention mechanism...
Specification
- Context window
- 164K
- Max output
- 66K
- Knowledge cutoff
- Not stated
- Parameters
- 685.4B
- Licence
- mit
- Serving providers
- 14
- Moderated
- No
- Uptime
- 100.0%
Intelligence
25.1
68th percentile
Coding
—
Coding Index
Agentic
—
Agentic Index
Output speed
100 t/s
Median across providers
Latency
676ms
Time to first token
Cost per task
$0.05
Estimated
Benchmarks
Where the score comes from
The Intelligence Index is a composite. These are the underlying evaluations this model was actually measured on.
Evaluation scores
Percentage correct · higher is better
- MMLU-Pro83.7%
- τ²-bench (Telecom)78.9%
- GPQA Diamond75.1%
- LiveCodeBench59.3%
- AIME 202559.0%
- IFBench49.0%
- AA-LCR (long context)42.7%
- SciCode38.7%
- Terminal-Bench Hard32.6%
- Humanity's Last Exam11.2%
An evaluation missing from this list was not run for this model — it is not a zero.
View as table
| Evaluation | Score |
|---|---|
| MMLU-Pro | 83.7% |
| τ²-bench (Telecom) | 78.9% |
| GPQA Diamond | 75.1% |
| LiveCodeBench | 59.3% |
| AIME 2025 | 59.0% |
| IFBench | 49.0% |
| AA-LCR (long context) | 42.7% |
| SciCode | 38.7% |
| Terminal-Bench Hard | 32.6% |
| Humanity's Last Exam | 11.2% |
Against its peers
Intelligence Index · this model highlighted, nearest peers in grey
- Claude Opus 4Anthropic26.0
- Gemini 2.5 ProGoogle25.9
- GPT-5 MiniOpenAI25.8
- Gemini 3.1 Flash LiteGoogle25.6
- 25.6
- V3.2DeepSeek25.1
- Qwen3 MaxQwen24.5
- Ling 3.0 Tiny (free)InclusionAI24.3
Peers are the models sitting closest on the Intelligence Index — the set you would realistically choose between.
View as table
| Model | Intelligence |
|---|---|
| Claude Opus 4 | 26.0 |
| Gemini 2.5 Pro | 25.9 |
| GPT-5 Mini | 25.8 |
| Gemini 3.1 Flash Lite | 25.6 |
| Gemini 3.1 Flash Lite Preview | 25.6 |
| V3.2 | 25.1 |
| Qwen3 Max | 24.5 |
| Ling 3.0 Tiny (free) | 24.3 |
Percentile among all indexed models
Pricing
What it costs to run
List prices per million tokens, plus what one representative task works out to.
List price
- Input / 1M tokens
- $0.269
- Output / 1M tokens
- $0.4
- Cached input / 1M
- $0.135
- Blended 3:1
- $0.302
One task, estimated
$0.05
- Input tokens
- 50,000
- Output tokens
- 80,000
- Profile
- Reasoning
Estimated from list pricing: 50K input tokens plus 80K output tokens for reasoning models (25K for non-reasoning).
Serving providers
- Baidu
- StreamLake
- DigitalOcean
- SiliconFlow
- DeepInfra
- AtlasCloud
- Novita
- GMICloud
- Venice
- Alibaba
- Friendli
- Phala
- SambaNova
Speed and latency figures are medians across these providers, so a widely-served model reports a blend rather than any single endpoint.