Instant, unlimited hosting for any Llama model on HuggingFace.
No servers needed.

Featherless is an LLM hosting provider that enables the use of a continually expanding library of models via the simplest possible API.

Forget about servers or per-token pricing and just use models via API.

Our model catalog is the largest of any single provider on the internet (by over 10x).

We have plans for both individuals and businesses.

Individual plans cost $25 per month for any model, or $10 for small models only (up to 15B).

Business plans start at $75 per month and scale to hundreds of concurrent connections.

No. We do not log your chats - not the prompts, not the completions.

We do log meta data - e.g. which models are used, what prompt lengths - as this is necessary to monitor and scale our infrastructure.

Please see our privacy policy for more details.

We support a wide range of model families - Llama2, Llama3, Mistral, Qwen, DeepSeek and RWKV.

Please see our model page for the complete list

Yes! Please reach out to us on Discord or email us

As we grow, we aim to automate this process to encompass all publicly available Hugging Face models with compatible architectures.

We price by concurrency, and not tokens, so your bill is predictable no matter how much you are using the service.
As long as you remain subscribed, there's no time cap on model usage.

Output is delivered at a speed of 10-40 tokens per second, depending on the model and prompt size.

Questions about model use and sampler settings are best asked in our Discord.

For account or billing issues, please email us.

Yes, we run all models at FP8 quantization to balance quality, cost and throughput speed.

We've found this quantization does not noticeably change model output quality, while significantly improving inference speeds.

At the heart of the platform is our custom inference stack, in which we can dynamically swap out models on the fly in <1 second for a 10B model.

This allows us to rapidly reconfigure our infrastructure according to user workload and autoscale accordingly, as a single unified unit.

Cost, speed, and customization

While cloud GPU providers like RunPod allow you run any model, cost of the GPUs is significant (minimum $2/hour for GPUs to run a 70B model).

There are services that abstract the model management (e.g. Replicate or Open Router), but these offer a much more limited array of models.

Featherless offers the complete range of models with none of the complexity or cost of managing servers or GPUs.

Yes! Refer a friend, and when they subscribe and add your email, both of you get $10 OFF your next monthly bill!

Refer 12 of your friends and you can have a full year off our basic plan! (The discount stacks!)

Details here.

We are a serverless AI hosting provider. Our product simplifies the process of deploying models from Hugging Face. We are not charging for the underlying models.

We list all supported models, any of which can be made available for inference in milliseconds. We interpret all API requests as model allocation requests and "deploy" the underlying model automatically. This is analogous to how an individual would use RunPod, but at a different scale.

Moreover, we are in contact with model creators of the most popular models to avoid misunderstandings about the nature of Featherless, and have obtained their permission. If you are a model creator and take issue with a listing here, please contact us at hello@featherless.ai.