--- title: "Kimi K3.1 Pro weights are available, with a custom license for deployment" description: "Kimi K3.1 Pro weights are publicly listed. Examine the repository, custom license thresholds and practical checks before planning a self-hosted deployment." url: "https://promptszone.com/kimi-k3-1-pro-weights-available-custom-license" published: "2026-11-05T03:00:05Z" categories: ["New Models"] --- # Kimi K3.1 Pro weights are available, with a custom license for deployment Moonshot’s public model files turn a promised release into a deployment option. The custom license and infrastructure requirements need careful review. ![Original diagram of 96 file tiles beside the official Kimi family logo, illustrating K3.1 Pro public weights and its custom license.](https://promptszone.com/assets/uploads/b82ecace02614f05a151e6099f7f35c4.png) **Summary:** Kimi K3.1 Pro’s official repository now lists downloadable weights. What the files establish, which license conditions matter and what teams still need to test. **Open-weight models · Availability checked November 4, 2026 (America/Chicago)** Moonshot AI’s Kimi K3.1 Pro weights are now publicly listed in its official Hugging Face repository, giving teams a concrete self-hosting option to examine after the earlier hosted launch. The [file listing](https://huggingface.co/moonshotai/Kimi-K3.1-Pro/tree/main) shows a roughly 1.56 TB repository and weight files numbered as parts of a 96-shard set. This is an availability update to [our original K3.1 Pro launch coverage](https://promptszone.com/kimi-k3-1-pro-launch-open-weights-november), which reported Moonshot’s promise to release weights by November 4. We verified the public listing on November 4 in Chicago time; we have not established the exact moment it became public or downloaded the complete model. ## Deployment control comes with a large footprint The [official model card](https://huggingface.co/moonshotai/Kimi-K3.1-Pro) lists 2.8 trillion total parameters, 104 billion active parameters and a 1,048,576-token context window, with text and image inputs. It links deployment guidance for vLLM, SGLang and TokenSpeed. For infrastructure teams, the useful change is access to the artifacts needed to investigate deployment, rather than another promise of future access. The repository’s size is a storage observation, not a minimum GPU-memory specification. Usable capacity also depends on the serving engine, context length, concurrency and runtime overhead. We have not verified a hardware configuration or measured local throughput. ## Read the custom license before planning a service The weights use the [Kimi K3.1-Pro License](https://huggingface.co/moonshotai/Kimi-K3.1-Pro/blob/main/LICENSE), not an unrestricted MIT license. It grants broad use and modification rights subject to conditions, including retaining copyright and permission notices. - A licensee or affiliate operating a model-as-a-service business must obtain a separate Moonshot agreement before commercial use if their combined revenue exceeds US$20 million over any consecutive 12 months. - A commercial product or service exceeding 100 million monthly active users or US$20 million in monthly revenue must prominently display the model name. - The license exempts defined internal use and access through Moonshot’s official products or certified inference partners from those two requirements. These are different thresholds and should not be collapsed into a single “free for commercial use” label. The license’s definitions matter, particularly its distinction between model-as-a-service, embedded product features and request relaying. This is a reporting summary, not a determination that a particular deployment is compliant; organizations should review the full terms against their own use. ## Why this update does not establish a new performance lead Public weights are new deployment evidence, not a new benchmark result. Moonshot’s [evaluation table and footnotes](https://huggingface.co/moonshotai/Kimi-K3.1-Pro#3-evaluation-results) combine its own tests with results from release reports and benchmark leaderboards. Harnesses and reporting sources vary across the compared systems. We are therefore not turning those mixed-method results into a ranked scorecard for this availability update. A useful independent comparison would pin model versions, harnesses, reasoning settings and task budgets, then run the same tasks under matched conditions. Readers can inspect Moonshot’s detailed results, but the new evidence here concerns access to the weights and their license. ## What to test before committing **Our evaluation checklist:** pin the repository revision, record the engine version and have the infrastructure team size a small pilot before downloading into production. Compare completed-task quality, latency and operating cost against the hosted option using the same private test set. Include failed tasks, retries and human review time rather than counting only successful demonstrations. Keep deployment and capability decisions separate. Public weights can expand operational choices, but they do not establish that self-hosting is cheaper, that every advertised context length is practical on a chosen cluster, or that the model is the strongest fit for a particular workload. Those conclusions need measurements under the intended conditions. Reporting: official repository, model card and license checked November 4, 2026, America/Chicago. No independent benchmark or deployment test performed. Cover: original PromptsZone vector graphic showing 96 schematic file tiles, with the unmodified official Kimi family logo used as the model-logo fallback. Logo: Moonshot AI, from its [brand guide](https://moonshotai.github.io/Branding-Guide/), used under its [editorial logo terms](https://www.kimi.ai/policies/logo-usage-terms). No affiliation or endorsement. Not a product screenshot.