TLDR: Chinese AI startup DeepSeek has released DeepSeek-V3.2-Exp, an experimental version of its foundation model. This new model incorporates DeepSeek Sparse Attention (DSA) technology, aiming for significant improvements in training efficiency and long-context processing. Described as an ‘intermediate step’ towards their next-generation architecture, the release highlights DeepSeek’s focus on reducing computational costs and challenging rivals in the AI landscape.
BEIJING – DeepSeek, a prominent Chinese artificial intelligence developer based in Hangzhou, announced on September 29, 2025, the release of an experimental version of its latest foundation model, DeepSeek-V3.2-Exp. The company characterizes this new offering as a crucial ‘intermediate step toward our next-generation architecture,’ signaling ongoing advancements in its AI research and development.
The DeepSeek-V3.2-Exp model introduces a groundbreaking innovation: DeepSeek Sparse Attention (DSA) technology. This mechanism is specifically designed to explore and validate optimizations for training and inference efficiency, particularly in scenarios involving long text sequences. According to DeepSeek, DSA achieves ‘fine-grained sparse attention with minimal impact on output quality’ while simultaneously boosting long-context performance and significantly reducing computational requirements.
This experimental release underscores DeepSeek’s commitment to developing more efficient transformer architectures. The company stated on the Hugging Face developer forum that this release ‘represents our ongoing research into more efficient transformer architectures, particularly focusing on improving computational efficiency when processing extended text sequences.’ The benefits of this new technology include more efficient training and an enhanced ability to process lengthy text inputs compared to previous iterations.
DeepSeek has previously made waves in the AI community with its V3 and R1 models, which reportedly delivered performance comparable to leading models from industry giants like OpenAI and Google, but at a fraction of the cost. The introduction of DeepSeek-V3.2-Exp, with its focus on cost reduction and efficiency, further solidifies the company’s competitive stance. DeepSeek also announced plans to cut its API prices by 50%, a move that could intensify competition in the global AI market.
While the new DeepSeek model is not expected to launch in global markets anytime soon, its advancements could exert considerable pressure on Chinese rivals, such as Alibaba’s Qwen. Alibaba recently upgraded its Qwen platform with the Qwen 3 Max model, which reportedly surpassed OpenAI’s GPT-5 chat in LMArena’s benchmark earlier this month. DeepSeek’s V3.2-Exp model, by dramatically reducing computing costs while maintaining high performance, positions the company as a formidable contender against both domestic and international AI players.
Also Read:
- Novel Training Approaches for Diffusion Models Significantly Enhance Generative AI Efficiency
- Gemini 2.5 Flash-Lite Preview Crowned Fastest Proprietary Model with 50% Output Token Reduction
To facilitate comparative testing and allow users to evaluate the benefits of sparse attention firsthand, DeepSeek has temporarily retained API access for its V3.1-Terminus model until October 15, 2025. This strategic release highlights a fundamental advancement in AI efficiency, potentially lowering the computational barriers that have historically limited access to cutting-edge AI capabilities and enabling broader deployment of sophisticated language models.
