China’s Moonshot AI unveiled Kimi K3 on July 16, 2026, a 2.8-trillion-parameter open-weight model that narrows the gap between Chinese and Western frontier AI. The model went live via Kimi’s apps and API immediately. Full open weights release July 27.

On independent benchmarks, Kimi K3 lands fourth globally—behind Claude Fable 5 and GPT-5.6 Sol, but ahead of Claude Opus 4.8. That’s a meaningful position. A year ago, Chinese models trailed by larger margins. Now they’re competitive.

What Makes K3 Different

Kimi K3 is a mixture-of-experts model with native multimodal understanding. It processes images and text. Unlike its K2 predecessors, K3 ships with reasoning always on—meaning it thinks through problems step by step without user intervention.

The 1-million-token context window is substantial. Researchers can feed entire codebases, academic papers, or long documents into a single request. That changes how people use AI for research and coding.

Pricing and Access

Moonshot set aggressive pricing: $3 per million input tokens, $15 per million output tokens. That undercuts most Western models. OpenAI’s GPT-5.6 Sol costs more. Anthropic’s Claude Fable 5 costs significantly more. Price matters for developers building applications.

The open-weights release on July 27 means developers can run K3 on their own infrastructure, without API costs. That’s leverage against proprietary models.

The Competitive Landscape Shifts

Moonshot is reclaiming market position after a period of relative quiet. The company raised over $1 billion and is investing heavily in training. Kimi K3 signals China’s AI race is producing competitive models, not just interesting research.

Kimi K3 lands between Anthropic and OpenAI on performance. If you’re building AI applications, testing K3 this week costs nothing.