What it does
Kimi K3 is Moonshot AI's most capable model. Its launch post describes a 2.8-trillion-parameter model with native vision and a 1-million-token context window, calls it the first open model in the 3-trillion-parameter class, and says the full weights were due to be released by July 27, 2026. It is aimed at long-horizon coding, knowledge work and reasoning, and it uses maximum thinking effort by default, with low and high modes to follow. The post says the model can sustain long engineering sessions, navigate large repositories and orchestrate terminal tools. Its examples cover GPU kernel optimization, compiler development, game creation, chip design, research reports and video editing. It also says overall performance still trails the most powerful proprietary models. K3 is available in the Kimi apps, the Kimi Work desktop app, Kimi Code in the terminal and the Kimi API. The post lists two limits: generation quality can become unstable if an agent harness does not pass back the model's thinking history, and the model may make unexpected decisions on the user's behalf when intent is ambiguous, so it advises explicit constraints in the system prompt.
Capabilities
- 2.8T parameters with 16 of 896 experts active
- Native vision and a 1M-token context window
- Maximum thinking effort by default
- Available in Kimi apps, Kimi Work, Kimi Code and the API
- Described as an open model, with weights announced for release by July 27, 2026
When to use it
- You need a long-context model for work across large repositories
- You want a model for long-running software engineering tasks
- You want coding combined with reasoning over screenshots and other visuals
- You are producing research reports or other knowledge work with agents
