Moonshot AI Teases Kimi-K3, Billed as First Open 3T-Parameter Frontier Model

Moonshot AI has put up a countdown page on Hugging Face for Kimi-K3, its next open-weights model, timed to go live July 27, 2026. As of the snapshot, the weights are not yet public — roughly 2,800 people have signed up for a release notification — so everything on the page is vendor framing rather than measured performance. The headline claim is that K3 will be the first openly released model in the ‘3T class,’ positioning it against closed frontier systems on parameter scale alone.

The teaser lists four expected capabilities: a new architecture built on ‘Kimi Delta Attention’ and ‘Attention Residuals,’ native agentic behavior (tool calling, browsing, and multi-step planning), a long context window aimed at repository-scale code comprehension, and fully open weights hosted on the release page. Moonshot is pitching the model squarely at long-horizon coding, knowledge work, and reasoning — the same territory where agentic, code-focused systems are currently competing hardest.

The significance, if the release lands as described, is another push to close the gap between open weights and proprietary frontier models, particularly for developers who want a self-hostable option for agentic coding. The caveats are equally clear: there are no benchmarks, no independent evaluations, and no confirmed licensing or hardware requirements yet. A 3-trillion-parameter open model also raises immediate practical questions about who can actually run it, which won’t be answerable until the weights and technical details are out.