Moonshot AI has released the weights of Kimi K3.1 Pro, adding a self-hosting route after the model’s earlier API debut. The November 4 release changes how developers and researchers can access the system, rather than introducing a different version of the model.
The official repository describes a natively multimodal mixture-of-experts model with 2.8 trillion total parameters and 104 billion active parameters. It accepts text and images and lists a context window of 1,048,576 tokens. The model card documents deployment routes through inference engines including vLLM and SGLang.
Open weights with model-specific terms
The release uses the Kimi K3.1 Pro License. Teams should review those terms for their intended use and distribution rather than assuming the repository carries Apache or MIT licensing.
The card also documents hosted API access, so downloading the model is an additional option. Its scale makes this a substantial infrastructure project; the active parameter count does not represent the full amount of weight storage needed for deployment.
What self-hosting changes
Our practical reading is that weight availability gives operators more control over where inference runs and how the model is integrated. It also transfers responsibility for capacity, upgrades, monitoring and service reliability. That tradeoff is useful to evaluate explicitly before replacing a managed endpoint.
A team can begin by reproducing a small set of tasks it already runs through the hosted service. Keep the inputs and acceptance criteria fixed, then compare the local checkpoint’s outputs, latency and memory use under realistic concurrency. A successful single request is only the first part of that assessment.
Agent integrations require particular attention to conversation state. The model card says applications should preserve the complete assistant messages needed for multi-turn work, including the provider’s reasoning and tool-call fields. A generic client that silently removes required fields can produce misleading results about model quality.
Long context deserves an equally practical test. A large window may accommodate a repository or document collection, but the model still needs to use the relevant evidence correctly. Include tasks with contradictory or superseded information and check whether the final action follows the current requirement.
The weight release is significant because it makes those experiments possible outside Moonshot’s own serving environment. For organizations evaluating deployment control, it is a concrete new option, accompanied by the engineering and licensing work that operating a large model requires.