Kimi K3 Arrives: A 2.8-Trillion-Parameter, One-Million-Token Open Model
Chinese AI startup Moonshot AI announced Kimi K3 on July 17, 2026. The new model has 2.8 trillion parameters, a one-million-token context window, and native multimodal support for images. It is designed for long coding sessions, knowledge work, and deep reasoning.
The scale is striking, but size is not the only point. Moonshot calls Kimi K3 the world’s first “open 3T-class model.” It is an attempt to show how close an open model can move toward top closed systems such as Claude Fable 5 and GPT-5.6 Sol.
A 2.8-trillion-parameter open 3T-class model
Kimi K3 contains 2.8 trillion total parameters, up from one trillion in Kimi K2 and close to the three-trillion range. It does not use all 2.8 trillion for every answer.
The model uses a Mixture of Experts architecture. Of its 896 experts, effectively 16 are active during processing. It resembles a huge company that selects a small specialist team for each job, aiming to combine scale with efficiency.
Moonshot also introduced Kimi Delta Attention and Attention Residuals, claiming roughly 2.5 times better overall scaling efficiency than Kimi K2.
One million tokens and visual understanding for long-running work
Kimi K3’s one-million-token context window can hold long documents and large codebases. Native image understanding supports a “vision in the loop” workflow where the model writes code, inspects a screenshot, and then revises the visual result.
Moonshot emphasizes long-running agent work such as:
- Reading a large repository and continuing development through a terminal
- Improving games, front-end interfaces, or CAD work while inspecting images
- Reading many papers and equations, then creating research code and visualizations
- Turning research into tables, slides, and interactive reports
- Understanding source footage and producing edits or motion graphics
It is designed less as a chatbot for short questions and more as a model that can work with a large amount of material and tools for hours or days.
Has it surpassed Fable 5 or GPT-5.6 Sol?
Kimi K3 reports very strong results on some coding evaluations. Moonshot also acknowledges that the overall experience still trails Claude Fable 5 and GPT-5.6 Sol.
Benchmark results depend on more than the model. Kimi Code, Claude Code, and Codex provide different tools, reasoning time, and ways of preserving work history. Vendor evaluations are useful evidence of capability, but they should not be treated as an absolute ranking.
You can use the service now, but weights were scheduled for July 27
Kimi K3 is available through Kimi.com, Kimi Work, Kimi Code, and the Kimi API. The API model name is kimi-k3. At launch, the maximum reasoning setting was the default, with lighter low and high settings planned for later.
The word “open” needs qualification. At announcement time, the hosted service and API were available, but the full model weights were scheduled for release by July 27, 2026. At the time this article was published, users could not yet download and run the complete model themselves.
A 2.8-trillion-parameter model also does not run casually on a single Mac. Moonshot recommends a serving configuration with at least 64 accelerators. “Open model” and “model that runs on a home computer” are separate ideas.
Kimi K3 raises the ceiling for open models
Kimi K3 matters not because it immediately replaces Fable 5 or GPT-5.6 Sol, but because an open model now combines near-frontier scale, coding, visual understanding, and long-running agent work.
If the weights and technical report are released as planned, researchers and developers can inspect the system and integrate it into other services and tools. The real evaluation begins there. Beyond the headline size, watch whether the results can be reproduced and whether the model can be made efficient enough to use in practical environments.
Reference: Moonshot AI: Kimi K3—Open Frontier Intelligence. Specifications, pricing, and the release schedule reflect information available on July 18, 2026.
