Featherless has made Moonshot AI's Kimi K3 model available through its hosted inference platform. Developers can call the model through the same OpenAI-compatible chat completions interface used for other Featherless models, using the model identifier moonshotai/Kimi-K3.

The availability announcement is separate from Moonshot AI's original Kimi K3 release. It matters for Featherless users because it turns the open-weight model into a managed endpoint without requiring them to provision and operate the unusually large model themselves.

What is available on Featherless

Featherless says its current Kimi K3 deployment supports a 32K context window, FP8 quantisation and image input. The provider is working to expand the available context, so the hosted configuration should not be assumed to expose every capability or maximum context length described for the underlying model.

Vision support means a request can combine text with an image URL in the familiar chat-completions message format. That can support tasks such as interpreting screenshots, documents or visual content, subject to the model's accuracy and the application's data-handling requirements.

FP8 reduces the numerical precision used for inference and can make a model of this scale more practical to serve. It can improve memory efficiency and throughput, although output quality and performance still need to be tested against each workload. Featherless has not published independent latency or throughput results for this deployment in the announcement.

A very large mixture-of-experts model

Kimi K3 is an open-weight mixture-of-experts model from Moonshot AI. Featherless describes it as having about 2.8 trillion total parameters and roughly 104 billion active parameters for each token, selecting 16 experts from a pool of 896. In a mixture-of-experts design, only part of the full model is activated for a given token, which can reduce the compute required compared with using every parameter on every step.

The model also uses Moonshot's Kimi Delta Attention architecture and a new MoonViT-V2 vision encoder. Featherless repeats Moonshot's claim that the attention design delivers substantially better scaling efficiency than Kimi K2. That is a developer claim about the architecture rather than an independently verified performance guarantee.

Featherless highlights benchmark results published by Moonshot for coding and scientific reasoning. Benchmark scores can help place a model, but they do not predict performance on every production task. Data contamination, prompting, evaluation settings and serving precision can all affect results. Teams should reproduce relevant evaluations using the hosted FP8 configuration and their own representative data.

Operational and governance considerations

A managed endpoint removes much of the infrastructure work involved in serving a multi-trillion-parameter model, but it adds provider-specific considerations. Buyers should measure request latency, rate limits, regional availability, failure behaviour and total usage cost before relying on the service for a production workflow.

Featherless says it does not log prompts or completions. Organisations handling personal, confidential or regulated information should still review the applicable privacy terms, retention controls, subprocessors and contractual commitments rather than relying on a short launch statement alone.

The current 32K context limit is another important boundary. Applications designed around a larger native context may need chunking, retrieval or summarisation when they move to the Featherless endpoint. Developers should also test vision inputs for unsupported formats, large files and ambiguous or adversarial content.

Why the availability update matters

Open weights make it possible to inspect, adapt and independently host a model, but the hardware and engineering required for Kimi K3 remain substantial. Hosted access gives smaller teams a faster way to evaluate it and compare it with other models behind a familiar API.

For Featherless, adding Kimi K3 broadens its catalogue with a frontier-scale multimodal option. The practical value will depend on how the provider expands context support and how the hosted configuration performs on real workloads. The model is available now, while larger-scale and customised deployments are offered separately through Featherless's dedicated GPU service.