The open-weight race now has a 2.8T-parameter entry. Moonshot AI’s Kimi K3 combines a 1M-token context window, native visual understanding, and long-horizon coding support in a model the company frames as the first open model in the 3T class. The real stake is not another benchmark table; it is whether open-weight systems can keep pressing into the coding and knowledge-work territory dominated by closed frontier models.

In the Kimi K3 technical blog, Moonshot says the model is built on Kimi Delta Attention and Attention Residuals, with a Mixture of Experts design that activates 16 of 896 experts. The company says those architectural choices, plus training and data changes, give Kimi K3 about 2.5x the overall scaling efficiency of Kimi K2. Those are first-party claims, but the scale and architecture are enough to make this one of the more consequential open-weight releases of the week.

Availability matters here. Kimi K3 is already accessible on Kimi.com, Kimi Work, Kimi Code, and the Kimi API, while the full model weights are scheduled for release by July 27, 2026. That creates a two-step rollout: developers can test the model through hosted products now, then evaluate self-hosting, customization, and third-party inference once the weights and technical report arrive. Moonshot says it is coordinating with inference partners and open-source maintainers before the broader release.

The strongest product angle is coding. Moonshot says Kimi K3 can work through large codebases, coordinate terminal tools, and use screenshots or visual feedback for game development, frontend engineering, and CAD-like workflows. Its case studies include GPU kernel optimization, a compact Triton-like compiler called MiniTriton, a 48-hour autonomous chip-design experiment, and a computational astrophysics reproduction that reviewed more than 20 papers and generated over 3,000 lines of Python.

The careful read is not that Kimi K3 has dethroned every proprietary model. Moonshot explicitly says its overall performance still trails Claude Fable 5 and GPT 5.6 Sol. The pressure comes from the package: 2.8T parameters, 1M context, native multimodality, and full weights on the calendar. The next checks are whether independent benchmarks match the launch claims, how expensive inference becomes at this scale, and whether developers can reproduce the long-horizon coding gains outside Moonshot’s own demos.

#kimi-k3

Kimi K3 puts 2.8T open weights into the frontier-model race