Overview
- Moonshot released Kimi K3, a 2.8‑trillion‑parameter multimodal model with a one‑million‑token context window that the company says will have its full weights published on July 27.
- The model posted notable wins in early comparisons, topping Arena’s blind coding leaderboard and producing a reported H100 CUDA kernel that ran about 14.8 times faster than optimized PyTorch on the same task.
- Moonshot set low API prices for Kimi K3—for example $3 per million input tokens on cache misses and $15 per million output tokens—which the company says undercuts comparable U.S. flagship pricing.
- The launch rattled markets and accelerated finance moves: U.S. chip and AI stocks fell after the reveal and Moonshot is pushing for a Hong Kong IPO within six months after recent fundraising and rapid revenue growth.
- Independent validation is still pending and analysts note the model’s reported parity with top U.S. systems is based on selected benchmarks and company data, while export‑control workarounds and efficiency tricks are being credited for progress under chip limits.