Model Details
DeepSeek-V3.2
Quick answer: DeepSeek-V3.2 is a significant update in DeepSeek's V3 model series, improving on V3.1 with stronger coding and reasoning performance. It maintains the 685B MoE architecture and MIT license while delivering meaningfully higher scores on SWE-bench and hard reasoning benchmarks. Multiple variants — V3.2-Exp, V3.2-Speciale, and V3.2 (Thinking) — target different capability profiles.
What DeepSeek-V3.2 Is
DeepSeek-V3.2 represents a substantial step forward in the V3 line, incorporating training improvements accumulated since V3-0324 and V3.1. The .2 designation signals more than a checkpoint update — it includes architectural or training methodology refinements that produce measurable benchmark gains across coding (SWE-bench), science (GPQA), and math (AIME/MATH-500).
DeepSeek simultaneously released specialised variants: V3.2-Exp (experimental, early-access capabilities), V3.2-Speciale (task-specific optimisation), and V3.2 Thinking (extended chain-of-thought mode).
Specifications
| Field | Value |
|---|---|
| Organization | DeepSeek |
| Total parameters | 685 billion (MoE) |
| Active parameters per token | 37 billion |
| Context window | 128,000 tokens |
| License | MIT |
| Modality | Text |
Pricing
| Input (per 1M tokens) | Output (per 1M tokens) | |
|---|---|---|
| DeepSeek API | $0.27 | $1.10 |
Related Models & Benchmarks
- Compare: DeepSeek-V3.1 · DeepSeek-V3.2-Exp · DeepSeek-V3.2 Thinking
- Benchmarks: SWE-bench Verified
Last updated 2026-06-19.
DeepSeek-V3.2
DeepSeek-V3.2 is a large language model developed by DeepSeek.
DeepSeek-V3.2 vs Alternatives
| Model | BrowseComp | BrowseComp-ZH | HLE | License |
|---|---|---|---|---|
| DeepSeek-V3.2 | 51.4% | 65.0% | 40.8% | MIT |
| DeepSeek-V3.1 | — | — | — | MIT |
| DeepSeek-V4 Pro | — | — | — | Proprietary |
DeepSeek-V3.2 BrowseComp-ZH (65.0%) is substantially higher than English BrowseComp (51.4%) — strong Chinese-language web research. MIT license makes it a top open-source option.
Frequently Asked Questions
What is DeepSeek-V3.2?
DeepSeek's October 2025 model scoring 51.4% BrowseComp, 40.8% HLE. MIT — open-source Chinese and English search.Scores from Benchgen evaluations. Last updated 2026-07-24.