Hosted inference now provides another way to evaluate Kimi K3.1 Pro before committing to a self-hosted deployment.
Together AI added Kimi K3.1 Pro to serverless inference on November 5. Its model catalog lists moonshotai/Kimi-K3.1-Pro with a 1,048,576-token context window and NVFP4 precision. It supports image inputs alongside text, with function calling and structured outputs also listed.
Together’s published token rates
| Token category | Standard price per 1 million tokens |
|---|---|
| Input | US$2.00 |
| Output | US$10.00 |
| Cached input | US$0.20 |
For an illustrative request using 100,000 uncached input tokens and 10,000 output tokens, those rates imply US$0.30 in token charges. The calculation is US$0.20 for input plus US$0.10 for output. It excludes any other charges and is not a measured task cost. Accepted output and retries can change a job’s cost.
Post-training support has a separate serving limit
Together announced LoRA post-training support on November 4. The supported-models documentation lists B200-SXM resources, a 163,840-token training context and a 167,936-token generator context for this route. These limits differ from the serverless inference window.
Dedicated inference deployment of those adapters is not yet available in the documented path. A team can therefore evaluate post-training support without assuming it can immediately serve the resulting adapter through dedicated inference.
The public weights remain a separate deployment route
Moonshot’s Kimi K3.1 Pro repository is publicly accessible and ungated, replacing the pending-weights status in our earlier report. The repository metadata inspected on November 4 lists 96 safetensors weight shards. Its creation and last-modification dates are November 2; those metadata dates do not establish the exact time public access opened.
A public weights listing changes the evaluation path
The model card reports 2.8 trillion total parameters and 104 billion active parameters. It describes MXFP4 weights and MXFP8 activations. These are the publisher’s specifications, not measurements from a MustHave deployment.
We inspected the repository metadata and documentation. We did not download or run the weights. A team considering self-hosting still needs to establish a supported inference configuration, sufficient hardware, and acceptable operating costs.
The custom license contains commercial conditions
The Kimi K3.1-Pro License permits use and modification subject to its terms. Copies or substantial portions must retain the copyright and permission notices. The document also defines a Model as a Service business and attaches a separate-agreement requirement to certain commercial use.
The separate-agreement requirement applies when a licensee or its affiliates operate that business and exceed $20 million in aggregate revenue. The measurement period is any consecutive 12 months, and the agreement is required before commercial use. The defined service category includes exclusions, so read the threshold with that definition.
A separate provision requires prominent display of the Kimi K3.1-Pro name for commercial products or services exceeding 100 million monthly active users or $20 million in monthly revenue. The license exempts specified internal use and use through Moonshot’s official products or certified inference partners from those two requirements.
These thresholds use different measurement periods. Review the full license against the intended deployment before treating the repository as an unrestricted commercial hosting option.
Keep hosted and self-hosted results comparable
Public weights provide a new route alongside hosted access. For a useful comparison, run the same bounded task from the same repository state and record the exact model files, inference configuration, and available tools. Compare accepted changes and review time alongside speed and cost.
A hosted result does not establish what a self-hosted setup will produce with a different configuration. Our model and developer news coverage follows those access changes. The next decision is which deployment you can reliably operate and evaluate.



