The Vibe Father
The Kimi K3 model is Moonshot's 2.8-trillion-parameter flagship with native vision, a one-million-token context window, and a sparse mixture-of-experts design. It targets long-horizon coding, end-to-end knowledge work, visual creation, and reasoning.
K3 is much more than a larger K2 checkpoint. Moonshot's official Kimi K3 architecture overview introduces Kimi Delta Attention, Attention Residuals, a larger expert layout, and a Stable LatentMoE system. The company says those changes improve overall scaling efficiency by about 2.5 times compared with Kimi K2.
Kimi K3 model specifications
| Feature | Kimi K3 | Why it matters |
|---|---|---|
| Total parameters | 2.8 trillion | Frontier-scale model capacity |
| Expert layout | 16 of 896 experts active | Sparse compute for each token |
| Context window | Up to 1M tokens | Large repositories, documents, and long agent histories |
| Input | Native text and vision | Code, screenshots, images, and visual feedback loops |
| Reasoning at launch | Max effort | Deep work with higher latency and output cost |
| Access | Kimi products and Kimi API | See the Kimi K3 rollout timeline |
Kimi Delta Attention
Kimi Delta Attention is designed to improve how the model handles information across long sequences. Long context is not useful if a model loses important dependencies or spends too much compute attending to everything equally. KDA is part of Moonshot's effort to make the one-million-token window practical for real tasks.
The public launch report gives the architectural direction, not every implementation detail. More complete training and architecture information is planned with the technical report.
Attention Residuals
Attention Residuals change how information flows across model depth. Deep models can struggle to preserve useful signals as representations move through many layers. Moonshot presents AttnRes as a way to improve that flow and get more useful capability from the additional scale.
Stable LatentMoE and sparse experts
K3 routes each token through 16 experts from a pool of 896. This mixture-of-experts design allows the model to have enormous total capacity without using all 2.8 trillion parameters for every token. Stable LatentMoE is meant to make that routing and training more effective.
Sparse activation does not make K3 small. The full expert pool still has to be stored and served. It mainly changes the amount of computation required for each token.
Native vision in the coding loop
Kimi K3 can inspect images and video alongside text. For coding, the most interesting use is vision in the loop. The model can build an interface, inspect the rendered result, compare it with the target, and revise the code.
Moonshot highlights browser games, 3D scenes, motion graphics, dashboards, and slide creation. Our own Kimi K3 Minecraft-style test produced procedural terrain, animals, mining, building, water, flight, and a day-night cycle inside one browser project.
Long-horizon work
The launch examples include a compiler, chip design, scientific code, industry research, and multi-step knowledge work. These are chosen to show continuity across long tasks rather than one-shot code completion. The key question is whether independent evaluations reproduce that strength across different tools and prompts.
How strong is Kimi K3
Moonshot reports leading or near-leading results across several coding and agentic suites. Harnesses differ across some comparisons, and several results still need independent replication. That is why our Kimi K3 Vibe Bench profile keeps the evidence and confidence visible instead of treating one vendor chart as final.
Use our benchmark methodology to see how we normalize source data and label results that still need independent replication.
