DeepSeek V3.2 Exp
DeepSeek V3.2 Exp is a large language model made by DeepSeek, added to the catalog on 29 Sep 2025. It has a context window of 163,840 tokens and costs $0.27 per 1M input tokens and $0.41 per 1M output tokens. It accepts text and produces text. Typical uses: efficient long-context text analysis, research in transformer optimizations and exploring sparse attention mechanisms.
Price checked ยท source: OpenRouter
What DeepSeek V3.2 Exp is
DeepSeek V3.2 Exp, released by DeepSeek on September 29, 2025, is an experimental text-to-text model designed to explore efficient transformer architectures. It features a large context window of 163,840 tokens and introduces DeepSeek Sparse Attention (DSA) for improved efficiency in long-context scenarios. Users can control reasoning behavior with a simple boolean parameter. The model supports various parameters like frequency_penalty, max_tokens, and temperature, making it versatile for research in efficient transformer design. While it maintains performance comparable to its predecessor, V3.1, it focuses on architectural optimizations rather than raw task accuracy.
Key features
- Experimental text-to-text model
- DeepSeek Sparse Attention for efficiency
- 163,840 token context window
- Control reasoning with a boolean
- Supports diverse parameters
- Research-oriented design
Typical uses
- Efficient long-context text analysis
- Research in transformer optimizations
- Exploring sparse attention mechanisms
- Comparative studies with V3.1
DeepSeek V3.2 Exp pricing
US dollars, list price through OpenRouter.
| Item | Standard |
|---|---|
| Input, per 1M tokens | $0.27 |
| Output, per 1M tokens | $0.41 |
No price change since it was added on 29 Sep 2025.
| What a task costs | Cost |
|---|---|
| Answer a short question500 in / 300 out | $0.0003 |
| Summarise a 50-page PDF25,000 in / 800 out | $0.0071 |
| Review a 2,000-line source file30,000 in / 2,000 out | $0.0089 |
| Read a 300-page contract set150,000 in / 1,500 out | $0.041 |
DeepSeek V3.2 Exp specifications
| Model id | deepseek/deepseek-v3.2-exp |
|---|---|
| Version | deepseek-v3.2-exp |
| Context window | 163,840 tokensAbout 122,880 English words |
| Max output | 147,456 tokens |
| Knowledge cutoff | 2025-07-31 |
| Added to catalog | 29 Sep 2025 |
| Input types | text |
|---|---|
| Output types | text |
| Tool calling | Yes |
| Structured outputs | Yes |
| Reasoning effort control | No |
| Content moderation | Not applied by provider |
Sample outputs
The same prompts are given to every model, and the unedited output is shown with the date it was produced. Upend does not publish leaderboard scores it did not run.
Write a short essay
RUN20 Nov 2026Write 150 words on how AI might positively upend work, leisure and creativity
AI promises to fundamentally reshape human existence by liberating us from routine. In work, it will automate mundane tasks, shifting our focus to strategic thinking, complex problem-solving, and empathetic roles, fostering a more fulfilling and human-centric economy. For leisure, AI will act as a hyper-personalized concierge, curating unique entertainment, travel, and learning experiences tailored to our deepest interests, making free time more enriching and adventurous. Most profoundly, in creativity, AI will serve as a collaborative partner for everyone. It will democratize artistic and intellectual expression, helping individuals compose music, write novels, or design innovations regardless of their prior training. This symbiosis won't replace human creativity but amplify it, unlocking unprecedented waves of collective imagination and personal growth.