Core capabilities
- 1M-token context for long agent sessions
- Open weights for self-hosted deployment
- Agentic training for continuous multi-step execution
- Strong Chinese and multilingual performance
← Model indexMODEL PROFILE / kimi-k3
Moonshot AI / current
Kimi K3
Moonshot's Kimi K3, repeatedly ranked among the strongest open-weight models.
Context1M
ModalityText · Image
AccessWeights · API
ReasoningAgentic
Where it fits.
Combines a one-million-token context with agentic training aimed at long sequences of continuous execution. Kimi K3 is Moonshot AI's open-weight flagship with a 1M-token context and agentic post-training. Public trackers place it at or near the top of open-weight rankings, with a successor K3.1 scheduled for mid-October 2026 and a much larger K4 in training.
Best-fit workflows
- Long continuous agent runs
- Open-weight self-hosting
- Chinese and multilingual work
04 / OPERATIONSFrom evaluation
From evaluation
to production.
Deployment notes
Good candidate where open weights and long agent runs both matter. Pin a specific checkpoint, monitor the K3.1 migration path, and benchmark on your own evals rather than relying on leaderboard positions.
Watch before choosing
Kimi K3.1 and larger successors are imminent; plan for rapid snapshot churn. Rapid release cadence makes snapshot pinning essential; benchmark claims move quickly across aggregators.
05
Family & versions
- Kimi K3Open-weight flagship
- Kimi K3.1Successor, imminent
Specifications and positioning checked against provider documentation.
Open official source↗