The world's first open-source model in the 3-trillion-parameter class — built on Kimi Delta Attention with native visual understanding, a 1M-token context window, and long-horizon agentic coding.
🌎 Curated by Michael Leung — based in Australia
Kimi K3 is Kimi's most capable flagship model to date, with 2.8 trillion parameters. It is built on Kimi Delta Attention (KDA), a hybrid linear attention mechanism, and Attention Residuals, with native visual understanding and a 1M-token context window. It is designed for frontier intelligence scenarios including long-horizon coding, knowledge work and reasoning. [official quickstart]
2.8TPARAMETERS
The first open-source model to reach 2.8 trillion parameters. In 9 of the past 12 months (2025/07–2026/07), Kimi models have held the frontier of open-source model scale. Full weights are scheduled for release by 27 July 2026.
KDA+ ATTENTION RESIDUALS
KDA and Attention Residuals help information flow through longer sequences and deeper models. With the Stable LatentMoE framework, the model efficiently activates 16 of 896 experts — roughly 2.5× the overall scaling efficiency of K2.
1MTOKEN CONTEXT
A 1M-token context window with automatic context caching, plus native image and video understanding — K3 can use screenshots and visual feedback to improve real workflows.
Two headline capability areas called out in the official documentation.
With minimal human supervision, K3 can sustain long-running engineering tasks, understand and work with large codebases, and coordinate terminal tools. It excels where software engineering meets visual reasoning — game development, frontend engineering, CAD and related scenarios.
K3 advances end-to-end knowledge work, with consistent advantages across production-oriented agentic workflows in Kimi's internal evaluations — beyond public benchmarks.
Why I built this mini-site.
According to my research on the internet, this model is very good at creating 3D games. I used K3 swarm agents to build one — a playable 3D web game after just 2 hours of agent coding — and I believe K3 can beat Fable 5 when it comes to coding. My write-up is here.
The API is OpenAI-compatible — point the official SDK at
https://api.moonshot.ai/v1, set model="kimi-k3", and you are running.
Thinking is always on, streaming, vision, tools and 1M context caching included.
