DeepSeek-V3.2-Exp

// Overview

DeepSeek-V3.2-Exp is an experimental open-weight large language model from DeepSeek, released on September 29, 2025 as part of the DeepSeek V3 line. It is built directly on DeepSeek-V3.1-Terminus and positioned by DeepSeek as an intermediate step toward a next-generation architecture rather than a clean-sheet model.

Its headline change is DeepSeek Sparse Attention (DSA): a two-stage attention path where a lightweight "lightning indexer" scores tokens across the context and a fine-grained selector keeps only the most relevant ones for full attention. Layered on the existing 685B-parameter Mixture-of-Experts stack (about 37B active parameters, with Multi-head Latent Attention), DSA cuts the cost of long-context training and inference while DeepSeek reports output quality on par with V3.1-Terminus.

DeepSeek-V3.2-Exp is text-only and ships under the MIT license, with open weights, a technical report, and GPU kernels (TileLang and CUDA) published on GitHub and Hugging Face. At launch DeepSeek made it the default chat/reasoner endpoint and cut API prices by more than 50%.

// Benchmarks

1. MMLU-Pro85%
2. GPQA-Diamond79.9%
3. Humanity's Last Exam19.8%
4. LiveCodeBench74.1%
5. AIME 202589.3%
6. HMMT 202583.6%
7. Aider-Polyglot74.5%
8. BrowseComp40.1%
9. SimpleQA97.1%
10. SWE Verified67.8%
11. SWE-bench Multilingual57.9%
12. Terminal-bench37.7%

Scores on a 0–100 scale (25-point gridlines); higher is better. Each benchmark links to its published source.

// Pricing

Official DeepSeek API launch pricing (USD), more than 50% below the prior V3.1-Terminus tier. Third-party hosts list close equivalents (OpenRouter ~$0.27 in / $0.41 out).

Pricing source ↗

// Strengths

• Introduces DeepSeek Sparse Attention (DSA) for substantially cheaper long-context training and inference
• Maintains benchmark parity with V3.1-Terminus despite the efficiency changes (e.g. MMLU-Pro 85.0, SimpleQA 97.1)
• MIT-licensed open weights with published technical report and open GPU kernels (TileLang + CUDA)
• Aggressive API pricing — more than 50% cheaper than the prior V3.1-Terminus tier
• Strong math and agentic-search scores (AIME 2025 89.3, BrowseComp 40.1)

// Best for

• Long-context document, codebase, and transcript analysis where attention cost dominates
• Cost-sensitive production deployments needing a cheap, capable open-weight chat/reasoner model
• Self-hosting on-prem with open weights and open GPU kernels
• Reasoning, math, and competitive-coding tasks
• Agentic / tool-use and web-search workflows (BrowseComp, SWE-bench)

// How to access

// DeepSeek V3 — every version

The full lineage of the DeepSeek V3 line, newest first. Every version has its own page — click any to compare specs, benchmarks and pricing.