Moonshot AIKimi K3
Kimi K3 is an open-weight, multimodal chat model developed by Moonshot AI. It is a 2.8T-parameter Mixture-of-Experts model built on a novel Kimi Delta Attention and Attention Residuals architecture, with a 1,048,576-token context window. The model is designed for long-horizon tasks such as coding, knowledge work, and reasoning, and features native vision capabilities for processing text and images.
Input
Output
Context
1M
Max Output
-
Parameters
2779.9B
Technical Specifications
Input Modalities
Output Modalities
Capabilities
Loading capabilities…
Benchmarks
Estimated GPU requirements for Kimi K3
Estimates based on INT8 quantization at up to 32K context. A count above one assumes tensor parallelism across the cards. Actual requirements vary by framework and configuration.
More models by Moonshot AI
The creator's other models in the catalog, with their context, size and license where published.
Compare it, or keep browsing
Put this model beside its alternatives on the same evidence, or go back to the full catalog.
Frequently asked questions about Kimi K3
Answered from the entry's own fields: context, license, modalities, evidence, serving and the memory to self-host.
Kimi K3 advertises a context window of 1,048,576 tokens. The figure is the creator's published maximum; a given provider may serve less, and the gateway routes on what each provider actually serves.
Yes. Kimi K3 is an open-weight model released under the Other license, so the weights can be downloaded and self-hosted within that license's terms.
Kimi K3 accepts Text and Image and produces Text. The capabilities card on this page lists which API features each deployment honours, such as function calling and structured output, with the source each was checked against.
Kimi K3 has published results from Artificial Analysis, shown by suite in the benchmarks card above exactly as the publisher reported them. Scores are not combined across suites, and a suite that has not measured Kimi K3 is shown as not available rather than estimated.
Yes. Kimi K3 is served on the managed pool through the OpenAI-compatible endpoint as moonshot-kimi-k3, pinned by name or chosen by routing when it is the best fit for a request. The Try in Playground button opens it directly.
About 2.67 TB at INT8 for the weights and a default context, from the catalog's 2779.9B parameter count; FP16 needs roughly twice that, and long contexts or many concurrent requests add KV cache on top. The GPU section on this page lists cards that hold it, and the capacity planner sizes it for your context length and traffic.
Fields collected from public registries, provider APIs and benchmark publishers, each tagged with its source.
Last updated: Oct 16, 2026
Put intelligence in the middle.
One gateway in front of every model, with your policies applied and every decision on record. Start on the free plan with 5,000 requests a month, no card required.