Particle.news

Download on the App Store

DeepSeek Releases Experimental V3.2-Exp With 50%+ API Cost Reduction

The open release targets lower long‑context costs to advance DeepSeek’s push to align models with next‑generation Chinese chips.

Overview

  • DeepSeek published V3.2-Exp on Hugging Face and Alibaba-backed ModelScope, with downloads also available via its website and app.
  • The company says the build improves training and inference efficiency and cuts API costs by more than 50% compared with prior versions.
  • TechCrunch describes V3.2-Exp as engineered to reduce inference costs for long-context operations.
  • The drop follows a rapid cadence of releases, arriving a week after V3.1‑Terminus and about two months after V3.1, with the original V3 launched in December 2024.
  • DeepSeek’s permissively licensed models are seeing wide developer uptake even as the company faces bans on some government and corporate devices in multiple countries.