Moonshot AI

AI · Open-Source · Foundation Models

Launch

Moonshot AI open-sources Kimi K3, a 2.8T-parameter MoE model

Only 104 billion of its parameters activate per query, showing the frontier race now rewards efficient sparsity over raw size.

• Moonshot AI published full open weights for Kimi K3 on Hugging Face at 00:00 UTC on July 27, 2026, calling it the world's first open 3-trillion-parameter-class model.
• Moonshot AI is a Chinese AI lab whose Kimi models compete with OpenAI, Anthropic and Google on frontier coding, reasoning and agentic tasks.
• Kimi K3 is a mixture-of-experts model with 2.8T total parameters but only 104B active per token, drawn from 896 experts with 16 selected plus 2 shared.
• It uses new Kimi Delta Attention and a Stable LatentMoE framework that Moonshot claims delivers roughly 2.5x more intelligence per unit of compute than Kimi K2, alongside a 1-million-token context window and native text/image/video understanding.
• The full weights span roughly 1.4-1.63TB in MXFP4 quantization, requiring around 8 H100 GPUs or 2TB+ of VRAM to self-host.
• K3 first launched via Kimi.com and the Kimi API on July 17, 2026, and traffic exceeded Moonshot's expectations, prompting it to pause new consumer subscriptions before the weights followed.
• Benchmarks put K3 fourth on Artificial Analysis's Intelligence Index at 57 but first on Arena's Frontend Code evaluation at 1,679, ahead of Claude Fable 5.
• By open-sourcing a frontier-tier model plus its attention kernels and MoE infrastructure, Moonshot narrows the gap between closed and open AI and pressures hosted API pricing across the industry.