DeepSeek V3.2
Sparse Attention at Scale
DeepSeek's late-2025 refresh of V3. The big change is "sparse attention" — instead of every token comparing itself to every other token, each token only attends to the most relevant ones. This drops inference cost dramatically at long context, letting the model handle 163K-token inputs without quadratic blow-up.
Intelligence
Medium
Cost
Low
$0.30 in / $0.45 out
Context
164K
Up to 163,840 tokens
How are Intelligence, Speed & Cost bucketed?
Intelligence and Speed buckets are percentile ranks on
Artificial Analysis. Cost buckets are fixed dollar
thresholds keyed off output-token price ($/M out).
Intelligence
- Top 1%≤ 1%
- Top 5%≤ 5%
- Top 10%≤ 10%
- Good≤ 25%
- Medium≤ 50%
- Below avg> 50%
Speed
- Top 1%≥ 345 tok/s
- Top 5%≥ 237 tok/s
- Top 10%≥ 196 tok/s
- Good≥ 146 tok/s
- Medium≥ 90 tok/s
- Slow< 90 tok/s
Cost
- Freeopen weights · self-host
- Low< $1 / M out
- Moderate$1–5 / M out
- High≥ $5 / M out
Why it matters
Establishes sparse attention as a viable production path, not just a research idea. Combined with R1's thinking and V3's MoE, V3.2 is the third major architectural axis DeepSeek has industrialized in a single year.
Core Capabilities
Long Documents
Handles entire codebases, books, and multi-doc RAG.
Research
Foundational paper or scientific contribution.
Generative
Produces images, video, audio, or other media.
Agent Workflows
Built for tool use and autonomous tasks.
Context Window
164k tokens
≈ 126 pages
4k Chat 聊天
32k Long docs 长文档
128k This model 本模型
400k Multi-doc 多文档
1M Codebase 整个代码库
10M
Availability
API
Not available
Product / App
Not available
Open Source
Released
Enterprise
—
Pricing Model
Free / self-host
Open weights — pay only for compute.
Self-host Capability / Performance
Where this model sits relative to the middle 60% of models in the tree. All scores are 0–10 (higher is better).
Lower 20% Upper 80% This model
Context / memory
Context window size · log-scaled
6.0
9.0
6.4
Lower 20% 20th percentile — 20% of models score below this This model Where the current model lands Upper 80% 80th percentile — only 20% of models score above this
Percentile boundaries are computed across every model in the tree that reports the underlying benchmark for each capability.
What it feels like
- Refresh of V3 with reasoning capabilities folded in.
- 671B-param MoE (37B active per token) — the open-weights model that finally matched GPT-4 / Claude 3.5 Sonnet quality
- 82.6% HumanEval (coding) and 90.2% MATH-500 — outperformed GPT-4o, Claude 3.5 Sonnet, and Llama 3 on each
- Trained for $5.576M on 2.788M H800 GPU hours — 1/100th of comparable proprietary training budgets
Best use cases
- Self-hosted frontier-quality LLM for orgs that can't use proprietary APIs
- Math, coding, and general knowledge benchmarks at open-source price
- Fine-tuning and distillation into smaller production models
Tools to try
Not ideal for
- Edge / single-GPU deployments — the MoE size is large despite low active params
- Multimodal use cases — text-only
Model Evolution
deepseek-v is DeepSeek's language model family.