Overview
DeepSeek V3.2 is the open-weight model that reset price expectations for the whole industry. Released December 2025, it's a 671B-parameter mixture-of-experts model (37B active per token) with a 128K context window, and its headline innovation is DeepSeek Sparse Attention (DSA) — an efficient attention mechanism that roughly halved long-context API cost overnight while preserving quality.
Beyond efficiency, V3.2 is the first DeepSeek model designed around reasoning with tools and agents rather than treating tool use as an afterthought. It posts strong scores — ~67.8% SWE-bench Verified, ~89.3% AIME 2025, ~85% MMLU-Pro — at a fraction of frontier pricing (often under $0.30 per million tokens), making it the default value pick for cost-sensitive production workloads. It ships under a permissive licence with open weights on Hugging Face.
Key capabilities
- DeepSeek Sparse Attention halves long-context cost
- First DeepSeek built for reasoning-in-tool-use
- 671B MoE with 37B active parameters
- Open weights and permissive licence
At a glance
SWE-bench Verified
~67.8%
AIME 2025
~89.3%
Price
From ~$0.14 / 1M tokens
Pricing: Industry-low — often under $0.30 per million tokens.
Pros & cons
What we like
- Unbeatable price-per-quality
- Open weights, self-hostable
- Strong agentic and reasoning performance
Trade-offs
- Trails the absolute frontier on the hardest tasks
- Named in 2026 distillation allegations and facing scrutiny
The verdict
DeepSeek V3.2 is the value champion of 2026 — frontier-adjacent quality at prices nothing else matches, with open weights for self-hosting. If your spend is dominated by a single expensive model and you haven't benchmarked V3.2, you're almost certainly overpaying.
Best for: Cost-sensitive production workloads and self-hosting.
Frequently asked questions
Why is DeepSeek V3.2 so cheap?
Its mixture-of-experts design activates only 37B of 671B parameters per token, and DeepSeek Sparse Attention cuts long-context compute roughly in half — together driving API prices well below frontier models.
Can I self-host DeepSeek V3.2?
Yes. It ships with open weights on Hugging Face under a permissive licence, so you can run it on your own infrastructure.
