China strikes again: Kimi K3, the open-source model that catches up to US leaders — now at OpenAI/Claude prices
Chinese startup Moonshot AI launched Kimi K3 on July 16, 2026 — its new flagship and China's biggest AI punch since DeepSeek. It's a mixture-of-experts (MoE) model with 2.8 trillion parameters (about 50 billion active per token, across 896 experts), native vision (images and video), and a 1-million-token context window.
Two angles make it historic: it's Chinese and it will be open source. Moonshot bills it as the first open '3T-class' model in history — no lab, American or Chinese, had ever promised to release the weights of a model at this scale. It's already available via web and API, and the company committed to publishing the full weights by July 27, 2026 (its K2 family is already on Hugging Face under a Modified MIT license).
Under the hood it debuts Kimi Delta Attention (KDA), a hybrid linear attention that decodes up to 6.3x faster at million-token contexts, plus 'Attention Residuals', which Moonshot says improves training efficiency by 25%. Together, the architecture scales about 2.5x more efficiently than Kimi K2's.
The benchmarks: between Opus 4.8 and the leaders
On the independent Artificial Analysis index, K3 debuts #4 in the world: on par with Claude Opus 4.8 and behind only Claude Fable 5 and GPT-5.6 Sol — unprecedented for an open-weights model.
// Intelligence Index — Artificial Analysis (v4.1)
On terminal agents it virtually ties the leader, and on frontier coding it lands below Claude but above OpenAI:
// Terminal-Bench 2.1 — Terminal agents (%)
// FrontierSWE — Frontier coding (%)
And there are categories where it outright leads, especially agents that browse the web and research on their own:
// Where Kimi K3 leads (%, reported by Moonshot)
Figures self-reported by Moonshot AI (each lab uses its own agent harness) + the independent Artificial Analysis index.
The price: the era of dirt-cheap Chinese AI is over
The other big story is what it costs. Kimi K3 runs $3 input / $15 output per million tokens ($0.30 for cached input) — the most expensive model any Chinese lab has ever shipped, and a huge jump from Kimi K2.5's $0.60/$2.50. That puts it squarely in OpenAI and Anthropic's pricing league: the same as a Claude Sonnet ($3/$15), on par with GPT-5.6 Terra ($2.50/$15), and not far from Claude Opus 4.8 ($5/$25) or GPT-5.6 Sol ($5/$30). The message is clear: China is no longer competing on price — it's competing on raw capability.
Source: Moonshot AI (blog de Kimi)
What does this mean for you?
If you want near-frontier capability without marrying OpenAI or Anthropic, Kimi K3 is the most serious open alternative to date — it shines at agents that browse the web, research, and use tools, and the weights land on July 27 (though self-hosting takes dozens of GPUs). But it's no longer the Chinese bargain: at $3/$15 it costs as much as a Claude Sonnet, so compare on quality, not price. And if you handle sensitive data, remember the hosted version runs on servers in China.