DeepSeek Unveils New Model With Sparse Attention, Slashes API Costs
00:00/00:00
Listen to this article
1x
Subscribe now
Listen to the full version
Listen to the full version
A user interface message on the DeepSeek artificial intelligence app on a mobile phone on Feb. 28, 2025.
Chinese AI startup DeepSeek has unveiled a new experimental large language model and slashed prices for its services, intensifying the price war in China’s crowded AI market.
The company on Monday released DeepSeek-V3.2-Exp, featuring a proprietary “DeepSeek Sparse Attention” mechanism that it says boosts training and inference efficiency for long texts. The model is now available on DeepSeek’s app, website, and mini program.
You've accessed an article available only to subscribers
VIEW OPTIONS
Unlock exclusive discounts with a Caixin group subscription — ideal for teams and organizations.
Subscribe to both Caixin Global and The Wall Street Journal — for the price of one.
- RELATED
Share this article
Open WeChat and scan the QR code
- DIGEST HUB
Digest Hub
Back
Explore the story in 30 seconds
- DeepSeek launched DeepSeek-V3.2-Exp with “Sparse Attention” for efficient long-text processing and cut API prices by up to 75%.
- The model’s performance is similar to V3.1-Terminus; earlier upgrades included faster inference and reduced hallucinations by 45–50%.
- DeepSeek’s R1 model research, featured in Nature, showed LLMs can learn reasoning autonomously via reinforcement learning.
AI generated, for reference only
Who’s Who
- DeepSeek
- DeepSeek is a Chinese AI startup that has launched an experimental large language model, DeepSeek-V3.2-Exp. This new model incorporates a "DeepSeek Sparse Attention" mechanism to enhance efficiency with long texts. The company has also significantly cut its API prices. DeepSeek's research has been recognized internationally, with its DeepSeek-R1 model's scientific principles featured in the journal Nature.
AI generated, for reference only
What Happened When
- May 2025：
- DeepSeek upgraded its R1 model, significantly improving performance in complex reasoning and reducing the hallucination rate by 45%–50%.
- August 2025：
- DeepSeek launched its V3.1 model, featuring a mixed inference architecture for faster responses.
- September 18, 2025：
- The scientific principles behind DeepSeek-R1 were featured as a cover story in Nature journal.
- September 22, 2025：
- DeepSeek released the V3.1-Terminus update, optimizing model performance in code and search agent tasks.
- September 29, 2025：
- DeepSeek released DeepSeek-V3.2-Exp with a new sparse attention mechanism and announced significant API price cuts.
AI generated, for reference only
- PODCAST
- MOST POPULAR
Sign in with Google
Sign in with Facebook
Sign in with 财新