Story perspectives

Moonshot AI's 2.8T Kimi K3 Beats Claude, Sparks U.S. Warning

7/18/2026

Summary

Moonshot AI Dominates

Moonshot AI unveiled Kimi K3 on July 17, 2026, a 2.8-trillion-parameter open-weight model.
Arena.ai placed Kimi K3 ahead of Claude Fable 5 on a coding benchmark.
Vercel CEO Guillermo Rauch said Kimi K3 outscored all proprietary models on a web test.
U.S. officials warned Kimi K3 could pressure AI firms after Z.ai fell 28% and the Nasdaq 100 dropped 1.5%.

Summary

Moonshot Builds MiniTriton

Moonshot built MiniTriton, a Triton-like compiler, and benchmarked it against Nvidia’s Triton on an L20 Ada card.
K3 has a 1 million-token context window and activates only 16 of its 896 expert modules per token.
Moonshot used MXFP4 weights and MXFP8 activations for training and advises running K3 on supernodes of 64+ accelerators.
K3 autonomously designed a simulated 4 mm² inference chip, reaching 100 MHz timing and over 8,700 tokens per second decode.

History

Historical Context

In February 2019 OpenAI announced GPT-2, a 1.5-billion-parameter open-weight model whose staged release sparked debate over open-source AI. The controversy prompted rivals such as Google DeepMind’s AlphaCode and later Moonshot AI to develop larger open models. In July 2026 Moonshot launched Kimi K3, an open-weight system that outperforms proprietary rivals, echoing the regulatory and investment scrutiny first raised by GPT-2’s $50,000 training-cost replication.

1 of 3

Moonshot AI Dominates

Moonshot AI unveiled Kimi K3 on July 17, 2026, a 2.8-trillion-parameter open-weight model.
Arena.ai placed Kimi K3 ahead of Claude Fable 5 on a coding benchmark.
Vercel CEO Guillermo Rauch said Kimi K3 outscored all proprietary models on a web test.
U.S. officials warned Kimi K3 could pressure AI firms after Z.ai fell 28% and the Nasdaq 100 dropped 1.5%.

1 / 3