Moonshot AI unveiled Kimi K3, a 2.8 trillion parameter open-weight model the company calls the largest of its kind, with benchmarks placing it near frontier US systems.
Key Points:
- Moonshot's Kimi K3 carries 2.8 trillion parameters and a 1 million token context window, with full weights due Jul. 27.
- Independent testing scores it near Anthropic's Opus 4.8, behind Fable 5 and OpenAI's GPT-5.6 Sol.
- The model tops Arena's front-end coding board, marking China's sharpest open-source challenge yet.
Kimi K3 Launch Details
Moonshot released the model on Jul. 16 through its app and interface, and it scheduled the full open weights for Jul. 27 under a permissive license. The Beijing startup, backed by Alibaba, built Kimi K3 on a mixture-of-experts design that fires only 16 of its 896 experts for each token it handles.
A 1 million token context window lets the model read sprawling codebases and long documents in a single pass, a capability Moonshot aims at software engineering and knowledge work. Native support for text, images, and video rounds out what the company frames as its most capable release yet. That 2.8 trillion parameter scale puts it well ahead of DeepSeek's V4 Pro, which runs near 1.6 trillion, while paid access starts at $3 per million input tokens and $15 per million output.
Moonshot positioned the debut days before the World Artificial Intelligence Conference in Shanghai.
Also Read: Solana Reclaims $77 As $4.15B DEX Surge Tests Bullish Momentum
Benchmark Rankings Frontier
Independent testing from Artificial Analysis scored Kimi K3 at 57 on its intelligence index, a mark that sits close to Anthropic's Opus 4.8 while trailing Fable 5 and OpenAI's GPT-5.6 Sol. On a separate benchmark of real-world work tasks it placed third, ahead of Opus 4.8. Developers on Arena ranked it first for front-end coding, above every leading US model tested.
The rankings carry weight because open weights let developers download and adapt a system rather than rent it, a shift that can move a lab's edge from raw model quality toward cheaper hosting and wider reach. For builders chasing AI agents, a permissively licensed model at this scale clears a familiar hurdle.
Moonshot Comeback Story
Analysts still urged caution, since many figures come straight from the company and independent testing remains thin so soon after launch. A Moonshot executive likened the parameter count to neural connections in the brain, arguing that more of them let the model store and reason over greater knowledge. Testers, meanwhile, flagged a higher hallucination rate than the prior Kimi model.
The launch caps a striking turnaround for Moonshot, whose market position had eroded over the past 18 months as DeepSeek surged to the front of China's open-source scene. Its earlier Kimi K2 models kept the lab in the conversation before Kimi K3 pushed that ecosystem closer to the proprietary systems it long trailed.
Read Next: SpaceX Erases Over $800B As Stock Falls Below IPO Price