DeepSeek launches DeepSeek v3.2 experiment with high efficiency

DeepSeek has released a new DeepSeek v3.2, which is an experimental model that delivers high efficiency over the previous model.

This model has achieved fine-grained sparse attention with minimized impact on output quality, improving long-context performance and reducing compute cost.

It is built on DeepSeek V3.1-Terminus, and the inference cost of the new experimental model is way lower than that of V3.2-Experimental. On the other hand, the benchmarks of this new experimental build are on par with V3.1 Terminus.

The V3.2 Experimental API is now available with $0.028 cach hit and $0.28 cache miss for input tokens and $0.42 on output. New prices are effective immediately from today.