DeepSeek Unveils V3.2-Exp: A New AI Model for Long Contexts
Chinese AI startup DeepSeek has unveiled its latest language model, V3.2-Exp. This experimental model boasts an efficient attention architecture for long contexts, making it a significant addition to the AI landscape.
V3.2-Exp is now accessible via multiple platforms, including the web interface, mobile apps for iOS and Android, API, and model checkpoints on Hugging Face. In benchmark tests, it performs on par with its predecessor, V3.1-Terminus, with slight variations in individual tasks due to reasoning-heavy tasks resulting in shorter answers.
DeepSeek has collaborated with Chinese chip manufacturers, including Huawei, to develop and optimize its models. V3.2-Exp can run on AI chips from Chinese providers like Huawei Ascend and Cambricon, positioning DeepSeek for the future as China aims to reduce dependence on US-based AI chip manufacturers.
The core update is DeepSeek Sparse Attention (DSA), which selects only relevant parts of the context for cheaper inference with very long inputs. This innovation reduces API prices by over 50 percent, maintaining similar performance to its predecessor. V3.1-Terminus remains accessible via a temporary API until October 15, 2025, for comparison tests.
DeepSeek's V3.2-Exp, with its efficient attention architecture and reduced API prices, is set to make a significant impact in the AI sector. By leveraging Chinese AI chips and reducing costs, DeepSeek is well-positioned to challenge western providers and contribute to China's AI chip independence.
Read also:
- AI-Generated Humor Spreads on Gemini Nano Banana: Light-hearted Modifications Spark Concerns over User Privacy
- China is, unlike the United States, embracing technological progress rather than attempting to restrict it.
- Social media spat between Elon Musk and Sam Altman features their confrontation; discord revolving around business rivalry in relation to Apple
- Amazon Drones Collide With Crane in Arizona, Deliveries Paused
