DeepSeek V3.2

DeepSeek·Sep 2025reasoningopen weights · MIT

68.4

modhub Index

DeepSeek V3.2 introduced DeepSeek Sparse Attention (DSA), cutting long-context inference costs dramatically and letting the company halve API prices to $0.28/$0.42. A 685B-parameter MoE with 37B active, it interleaves thinking and tool use, and held its own against models many times its price through late 2025. V4 has since superseded it at the frontier, but V3.2 remains a superb budget pick and a popular self-hosting target.

Benchmark results

Where it shines

  • Sparse attention makes long contexts unusually cheap
  • Strong math and coding for near-commodity pricing
  • MIT-licensed weights with an active ecosystem

Alternatives to DeepSeek V3.2

Frequently asked questions

How much does the DeepSeek V3.2 API cost?
DeepSeek V3.2 costs $0.28 per million input tokens and $0.42 per million output tokens, with cached input at $0.03 per million. A workload of 10M input and 1.5M output tokens per month costs about $3.43.
What is the context window of DeepSeek V3.2?
DeepSeek V3.2 supports a context window of 128,000 tokens (128K), with up to 64K output tokens per response.
Is DeepSeek V3.2 open source?
Yes — DeepSeek V3.2 is an open-weights model released under the MIT license, so it can be downloaded and self-hosted.
What are the best alternatives to DeepSeek V3.2?
The closest alternatives by overall capability are OpenAI o3, Kimi K2 Thinking, GPT-5, Gemini 2.5 Pro. See the comparison pages for detailed head-to-head breakdowns.