Particle.news
Download on the App Store
3 ARTICLES
·
1h ago

DeepSeek Open-Sources Experimental V3.2-Exp With New Sparse Attention and API Price Cut

The experimental model debuts a fine-grained sparse attention design aimed at boosting long-context efficiency without degrading benchmark results.

Image

Overview

  • DeepSeek released and open-sourced DeepSeek-V3.2-Exp, making the model available across its app, web and API, with a paper on GitHub and model files on HuggingFace and ModelScope.
  • The update introduces DeepSeek Sparse Attention, which the company says significantly improves training and inference efficiency on long texts while maintaining output quality.
  • Training settings were aligned with V3.1-Terminus, and public benchmark performance is reported as roughly on par to enable fair comparison.
  • API pricing was reduced by more than 50% for developers, and access to the V3.1-Terminus endpoint is temporarily retained to support side-by-side testing.
  • Huawei’s Ascend platform announced same-day support with open inference code and new operator implementations, citing sub-2-second TTFT and sub-30-millisecond TPOT at 128K using vLLM and SGLang integrations.