The Hangzhou-based firm said DeepSeek-V3.2-Exp introduces DeepSeek Sparse Attention to cut compute and improve performance on long sequences. It also posted on X that it is reducing API prices by “50%+.”
A model card on Hugging Face lists the release under an MIT license and shows a configuration in the MoE family at 685 billion parameters, with benchmark parity claims versus V3.1. Weights are available for developers.
V3.2-Exp is positioned as a step between current and upcoming designs, following a year in which DeepSeek’s R1 and V3 models drew attention for competitive performance at lower cost. In mid-September, the company said R1’s training spend was $294,000 on 512 Nvidia H800s, intensifying focus on efficiency.
Earlier DeepSeek releases helped trigger a market rout in January as investors reassessed chip demand and AI cost curves. Even if V3.2-Exp is incremental, sustained price cuts could pressure rivals from Alibaba’s Qwen in China to OpenAI in the United States.
DeepSeek did not provide independent verification for the price cut or detailed third-party benchmarks for V3.2-Exp.
The release did not disclose training dataset composition, energy use/emissions, training duration, or hardware counts for V3.2-Exp beyond high-level claims.
I truly appreciate you spending your valuable time here. To help make this blog the best it can be, I'd love your feedback on this post. Let me know in the comments: How could this article be better? Was it clear? Did it have the right amount of detail? Did you notice any errors?
If you found it valuable, please consider sharing it.