If you mean Moonshot AI’s “Kimi K2.5,” yes: Moonshot calls it open-source and publishes the model weights. For a precise licensing and deployment decision, treat it as an open-weight model under a modified MIT license, not as a model with no usage conditions. The official repository names it Kimi K2.5; “Kimi 2.5” alone does not establish that a separate model is meant. [1]
What is available—and what does the license allow?
Moonshot’s GitHub repository documents Kimi K2.5, and its Hugging Face model card identifies the license as modified MIT and provides model files. The materials also include deployment guidance. This establishes that the model files are published; it does not mean the training data or complete training process is released. [1][2]
The license text grants permission to use, copy, modify, merge, publish, distribute, sublicense, and sell the “Software” and associated documentation. It requires the copyright notice and permission notice to be included in copies or substantial portions. Its additional condition applies if the Software or a derivative is used in a commercial product or service with more than 100 million monthly active users OR more than US$20 million in monthly revenue: the product or service must prominently display “Kimi K2.5” in its user interface. These are alternative triggers, not cumulative ones. [3]
The model card labels the release modified MIT, but the license text describes its scope as the Software and associated documentation; it does not separately enumerate every model artifact. If you plan to redistribute particular weights or other files, check the applicable terms for those specific artifacts. [2][3]
For production: self-host the weights or use an API?
| Choose | When it fits | Kimi K2.5 consideration |
|---|---|---|
| Self-host the weights | You need control over deployment or data handling, and can support the serving setup. | Moonshot’s deployment guide gives a vLLM example using a single-node setup with eight H200 GPUs. This is an example configuration, not a stated minimum; self-hosting can require substantial hardware and engineering effort. [4] |
| Use a hosted API | You prefer managed inference and a provider currently offers the exact model on terms that fit your workload. | Moonshot’s API platform retired kimi-k2.5 on August 31, 2026. Do not assume it is available through that platform; check any other provider’s current model identifier, capabilities, limits, and service terms before building against it. [5] |
Practical choice: If you specifically need Kimi K2.5, first check whether a third-party provider currently serves that exact model. Otherwise, self-hosting the published weights is the route described in Moonshot’s deployment materials, provided your team and infrastructure can support it. If you need a managed Kimi API rather than K2.5 specifically, assess currently listed models separately; do not assume the retired K2.5 model name will work. [4][5][6]
Bottom line: Kimi K2.5 is released with public model files under a modified MIT license. The license grants broad permissions, including commercial use, subject to notice requirements and a UI-branding condition at the stated scale thresholds. Moonshot’s API platform retired the model on August 31, 2026.
Sources
MoonshotAI/Kimi-K2.5: Open Visual Agentic Intelligence
moonshotai/Kimi-K2.5 model card