Particle.news

Moonshot Releases Kimi K3, a 2.8-Trillion-Parameter Open-Weight Model

Early benchmark wins with a planned July 27 public weights release are forcing U.S. policymakers to re-evaluate export controls and rules on model distillation

Overview

  • Moonshot unveiled Kimi K3 in mid-July, publishing API access immediately and scheduling full model weights to be downloadable on July 27.
  • The model holds about 2.7–2.8 trillion parameters, supports a 1 million‑token context window, and accepts multimodal input such as text and images.
  • Independent and company tests show K3 competitive with top U.S. systems on coding, long-horizon reasoning, and knowledge work, and it ranked first on Arena.ai’s front-end coding leaderboard in early evaluations.
  • Moonshot priced K3 at $15 per million output tokens, undercutting some U.S. frontier offerings and increasing commercial pressure on higher‑priced providers while still requiring costly hardware to run locally.
  • The release has shaken markets and revived policy debate in Washington over ‘distillation’—training from another model’s outputs—and the limits of export controls, while Chinese leaders frame open-weight releases as a cooperation and access strategy.