Moonshot AI, the Beijing-based startup backed by Alibaba, has released Kimi K3 — a 2.8-trillion-parameter open-weight model that is now the largest ever made publicly available, and one that benchmarks show performing neck-and-neck with the most powerful proprietary systems from OpenAI and Anthropic.

The release triggered a global tech rout on July 17, sending AI and semiconductor stocks sharply lower as investors digested the implications of a Chinese model competing at the absolute frontier despite US export restrictions on advanced chips. Bloomberg reported the surprise breakthrough from Moonshot "rippled through global markets."

Kimi K3 features a 1-million-token context window, native vision capabilities, and always-on reasoning mode. It activates just 16 of its 896 experts per token (roughly 1.8% of parameters per inference), making it efficient despite its massive scale — a 2.5x improvement in scaling efficiency over Kimi K2.

On the GDPval-AA v2 benchmark, K3 scored third overall — behind only Claude Fable 5 Max and GPT-5.6 Sol Max, and ahead of Claude Opus 4.8. On the AA-Briefcase agentic benchmark, it climbed to second place. Most strikingly, it achieved a state-of-the-art score on BrowseComp (91.2/100) for long-horizon information seeking and claimed the 1 spot on Arena.AI's Frontend Code Arena, surpassing Claude Fable 5.

In a demonstration of autonomous capability, Kimi K3 independently designed a physical chip over 48 hours — from architectural design through verification — using open-source EDA tools, producing a functional 4mm² design that closed timing at 100 MHz.

The model is priced at $3 per million input tokens and $15 per million output tokens, with cached inputs at just $0.30 per million — competitive with mid-tier Western offerings at near-frontier performance. Full model weights are scheduled for public release on July 27.

Architectural innovations include Kimi Delta Attention (a hybrid linear attention mechanism) and Attention Residuals — both published as open research. The training reportedly used Nvidia H200 GPUs alongside an unnamed alternative vendor's hardware, highlighting how Chinese labs are working around US chip curbs.

For the open-source community, the significance is clear: the performance gap between open-weight and proprietary models at the frontier has functionally closed. "Open source is no longer lagging six months behind Western closed-source models," one commentator noted. K3 represents a watershed moment — and the AI arms race just got a lot more crowded.

Sources