DeepSeek · open weights · updated July 2026

DeepSeek V3.2

Parameters
671B / 37B active
Context window
128K tokens
License
MIT
Modalities
text
API price (in/out per 1M)
$0.28 / $0.42
Released
2025-12-01

DeepSeek V3.2 was the model that introduced DeepSeek Sparse Attention (DSA) to production, cutting long-context inference cost dramatically while holding benchmark parity with V3.1-Terminus. Through early 2026 it was the workhorse open-weight model for coding platforms and hosting providers, scoring around 70% on SWE-bench Verified under an MIT license.

With the V4 family's release in April 2026 it moved from flagship to value tier, and it remains a well-understood, widely-hosted target with mature vLLM and SGLang support — often the pragmatic choice when V4 Pro is overkill.

DeepSeek V3.2 specifications

VendorDeepSeek (China)
ArchitectureMixture-of-Experts with DeepSeek Sparse Attention (DSA)
Total parameters671B
Active parameters37B
Context window128K tokens
Modalitiestext
LicenseMIT — weights free to download, self-host, fine-tune, and use commercially
Release date2025-12-01
WeightsHugging Face · GitHub

DeepSeek V3.2 API pricing

DirectionPrice per 1M tokens
Input$0.28
Output$0.42

First-party API rate at launch; check current DeepSeek pricing — the V4 family has since become the flagship tier. Verified July 2026.

Benchmarks and reported results

BenchmarkResultNote
SWE-bench Verified~70%as reported in late-2025 coverage

Running DeepSeek V3.2 locally

671B total / 37B active. A 4-bit quantization lands around 350–400GB — an 8×A100/H100 node or a large unified-memory cluster. Popular with hosting providers; heavy for home labs.

What DeepSeek V3.2 is best for

  • Proven, widely-hosted open-weight coding capability
  • Teams standardized on the V3-generation deployment stack
  • Budget API workloads on the DeepSeek platform

Frequently asked questions

Is DeepSeek V3.2 still worth using after V4?

For many production workloads, yes. V3.2 is cheaper, extremely well supported across inference providers, and its ~70% SWE-bench Verified score remains competitive. V4 wins when you need the 1M-token context or peak reasoning quality.

What license does DeepSeek V3.2 use?

MIT. Weights, including the sparse-attention variants, are free for commercial use, modification, and redistribution.