Moonshot AI has released Kimi K3, the world's first open 3-trillion-parameter-class model featuring a 1-million-token context window and native vision capabilities. Open weights drop July 27.

Moonshot AI Releases Kimi K3: The First Open 3-Parameter-Class Model Hits the Market

2.8 trillion parameters, 1 million token context, and pricing that leans heavily on cache hits.

Moonshot AI has confirmed Kimi K3, its latest flagship model released on July 16, 2026. It's the world's first open 3-trillion-parameter-class AI model, though the architecture actually runs on 2.8 trillion parameters. The company rounds up for marketing. You can grab the open weights starting July 27, 2026.

Moonshot AI is one of China's "Six AI Tigers," a group of six leading AI companies competing against American frontier labs. The other five are Z.ai, MiniMax, 01.AI, Baichuan, and StepFun. Moonshot was founded in March 2023 by Tsinghua University classmates Yang Zhilin, Zhou Xinyu, and Wu Yuxin. It's named after Pink Floyd's The Dark Side of the Moon, released 50 years prior to the company's founding. The name is poetic. The hardware requirements are not.

Kimi K3 features a Mixture of Experts architecture. It has 896 experts total, but only activates 16 at a time. That's extreme sparsity. The model boasts a 1-million-token context window. That's roughly 1,573 A4 pages of text. It includes native vision capabilities. Moonshot calls the attention mechanism "Kimi Delta Attention." There's also "Stable LatentMoE" to handle routing. On top of that, the model uses quantization-aware training from the SFT stage using MXFP4 weights.

Keep in mind that Kimi K3 is positioned for frontier intelligence scenarios. Moonshot says the model is built for long-horizon software engineering, deep knowledge work, and complex reasoning. The rapid iteration schedule is striking. K2.6 shipped in April. K3 arrives three months later. The pace suggests Moonshot is trying to close the gap with US incumbents before they pull away further.

Running K3 on Consumer Hardware

The open weights drop July 27. But a 2.8 trillion parameter model doesn't exactly play nice with a standard laptop. That's where projects like Colibri step in.

Head here to check the GitHub repository if you want to test the approach yourself. The tool focuses on slicing massive context windows and experts layers across consumer-grade hardware. If you're watching the open-weight scene, it's a practical answer to running frontier architecture without a data center budget.

Performance

Next, the performance numbers. Kimi K3 trails Claude Fable 5 and GPT 5.6 Sol on most benchmarks. However, it outperforms Claude Opus 4.8 and GPT 5.5. On SWE Marathon, a measure of long-horizon coding endurance, K3 scores 42.0. That beats Claude's 35.0. It ranks #4 on the Artificial Analysis Intelligence Index with a score of 57.

The model is also capable of some impressive autonomous work. In a 48-hour run, K3 designed a chip for a nano model using open-source EDA tools. The chip achieves 100 MHz timing closure. It also developed MiniTriton, a compact compiler. And it optimized GPU kernels across NVIDIA H200 hardware. Moonshot notes the chip design achieved 8,700+ tokens/s decode throughput. A chip built by a model, for a model.

Pricing

The pricing structure is aggressive on cache but steep on raw generation. Input costs $0.30 for a cache hit or $3.00 for a miss. Output is $15.00 per million tokens. Moonshot claims a cache hit rate above 90% on coding workloads. The blended price comes to roughly $2.31 per million tokens using a 7:2:1 ratio. If you're running coding agents that hit the cache often, the Mooncake architecture makes this a much more affordable option. It's not cheap for raw generation, though the performance gains might justify the spend.

There are limitations. K3 is sensitive to thinking history. Switching sessions or using incompatible agent harnesses can cause unstable generation. It may also make unexpected autonomous decisions during task execution. Moonshot recommends explicit behavioral constraints for production use. However, at the same time, the gap with Claude Fable 5 and GPT 5.6 Sol is still noticeable in user experience.

Moonshot AI has ~300 employees. Alibaba holds a 36% stake following a $1 billion funding round in February 2024. On top of that, the company is considering a Hong Kong IPO as of March 2026. The long wait for an IPO might be ending sooner than expected.

Kimi K3 is available through the Kimi.com chatbot, Kimi Work desktop agent, and the API at platform.kimi.ai. The open release gives the community a chance to test the frontier capabilities directly.