Particle.news
Get it on Google Play
Download on the App Store
18 ARTICLES
·
last wk.

Moonshot Unveils Kimi K3, a 2.7–2.8 Trillion‑Parameter Open‑Weight Model

Planned public release of K3’s weights on July 27 will let developers run and modify a frontier‑class system, shift costs toward inference compute, prompt U.S. policy scrutiny.

Founder of Moonshot AI Yang Zhilin delivers a speech at a conference of 2026 Zhongguancun forum, in Beijing, China March 25, 2026. REUTERS/Tingshu Wang/File Photo
© Chip Somodevilla
In its release, Moonshot claimed K3 performed “competitively” with Anthropic’s Fable 5, currently the most advanced AI model widely available on the market today.
Image

Overview

  • Moonshot publicly launched Kimi K3 on July 16 and opened an API and early access demos for developers, with a full weights release scheduled for July 27.
  • The company says K3 runs about 2.7–2.8 trillion total parameters, supports a 1 million‑token context window, and natively handles text and images for long, multi‑step workflows.
  • K3 uses a Mixture‑of‑Experts architecture and an ‘Agent Swarm’ design that coordinates many sub‑agents, letting the model activate only a subset of experts per query to save compute.
  • Early independent and developer‑voted tests such as Arena.ai placed K3 at or near frontier level for coding and knowledge tasks, but broader peer‑reviewed benchmarks and long‑run reliability checks are still pending.
  • Moonshot is selling API access at substantially lower advertised prices than top U.S. models and is reported to be seeking fresh funding at about a $31.5 billion valuation, a combination that tightens U.S.–China AI competition and shifts focus to inference compute and export‑control policy.