Particle.news
Download on the App Store

Moonshot’s Kimi K3 Narrows U.S. Lead With Open 2.8T Model

A scheduled July 27 public weight release could let developers run a frontier model at much lower cost, reshaping market assumptions about who controls top AI capability.

Overview

  • Moonshot unveiled Kimi K3 on July 17 as a 2.8 trillion-parameter, multimodal, open-weight model that supports a one-million-token context window and new sparse architecture choices.
  • Public third-party evaluations and leaderboards placed Kimi K3 at or near the top on several coding and agent tests, including a first-place score on Arena’s coding leaderboard.
  • Moonshot priced API access well below U.S. rivals and plans to release full model weights on July 27, which would let developers download and run the model themselves and lower ongoing token costs.
  • The announcement triggered sharp sell-offs in chip and AI-related stocks, accelerated Moonshot’s push toward a Hong Kong IPO and followed large private funding that lifted the company’s valuation into the tens of billions.
  • Major open questions remain because Moonshot’s capability claims are largely self-reported, Anthropic has accused some Chinese labs of improper 'distillation,' and high demand forced Moonshot to temporarily pause new sign-ups while it scales compute.