China’s DeepSeek has launched a new AI model called DeepSeek-V3.2-Exp. The new model comes with the new DeepSeek Sparse Attention technology, which the company says is “designed to explore and validate optimizations for training and inference efficiency in long-context scenarios.”
DeepSeek says that the new model is an “intermediate step toward our next-generation architecture.”
The company had shocked Silicon Valley earlier in the year with the global rollout of its V3 and R1 models, which delivered similar performance compared to the leading models from OpenAI and Google while being built at a fraction of the cost.
The new DeepSeek model is built on the older V3.1 model by introducing new technology to explore and optimize AI training and operation. DeepSeek says that the new model is intended to show its improved efficiency while processing long text sequences.
”This experimental release represents our ongoing research into more efficient transformer architectures, particularly focusing on improving computational efficiency when processing extended text sequences,” DeepSeek said on Hugging Face.
In a post on X (formerly Twitter), DeepSeek said that the DSA achieves “fine-grained sparse attention with minimal impact on output quality” while boosting long context performance and reducing compute.
The company also says that it is to launch in the global cutting the API prices by 50%.
The new AI model from DeepSeek isn’t expected to launch in the global markets anytime soon. However, DeepSeek’s AI could still put pressure on its Chinese rivals like Alibaba’s Qwen.
Alibaba had also given Qwen a big upgrade earlier in the month with the launch of the Qwen 3 Max model, which beat OpenAI’s GPT-5 chat on LMArena’s benchmark.
About the Author
