Particle.news
Download on the App Store

Moonshot Releases Open-Weights Kimi K2 Thinking, Claiming Wins on Agentic Benchmarks Over GPT-5 and Claude 4.5

The open release challenges the paid model business case by claiming top agentic performance.

Overview

  • Moonshot made Kimi K2 Thinking publicly available with open weights on Hugging Face, positioning it as an open-source reasoning model.
  • Published results report higher scores than GPT-5 and Claude Sonnet 4.5 on hard tests, including 44.9 on Humanity’s Last Exam and 60.2 on BrowseComp.
  • The model uses a Mixture-of-Experts design with roughly 1 trillion parameters and 384 experts, activating about 32 billion parameters at inference.
  • Moonshot’s documentation and demonstrations highlight long-horizon tool use up to 200–300 sequential calls, including solving a PhD-level math problem and generating complex apps from a single prompt.
  • Moonshot cites an estimated $4.6 million training cost reported to CNBC, raising commercial and geopolitical questions as independent verification and security reviews remain pending.