Subscribe
Sign in
Featherless AI — Run every 🦙 AI model & more from 🤗 huggingface
Featherless is a platform to use the very latest open source AI models from Hugging Face.
With hundreds of new models daily, you need dedicated tools to keep with the hype.
No matter your use-case, find and use the state of the art AI model with Featherless.
40
Replies
@manisha_hr_ Thanks!
Its OpenAI API - so if your devs were already using OpenAI styled API for AI, they should be able to try this instantly (as its the same API)
That's crazy that you offer this for $10-20; read about how you are swapping out models in almost real time, that's super cool! Thank you for making open source AI world a better place. As someone without the knowledge on how to deploy LLMs, this is a cool way to test out all kinds of OSAI quickly!
Congrats @eugene_cheah !
@tonyhanded Thanks! Hope you find a model you like for your use case, or just to play with =)
How does featherless tie into the novel models that you developed with recursal? Is it similar to OpenRouter? What are some of the top reasons to switch over to Featherless vs using other API's?
Hunter
@the_esc
Expect to see the Eagle and RWKV models on featherless shortly - we targeted the Llama-3 based models first since it's more widely known and has a very rich and varied set of fine-tunes, but we definitely believe that the future belongs to RWKV :)
We're certainly similar to OpenRouter in providing an enormous range of models servelessly. We aren't listing other providers and we will likely complement OpenRouter by acting as a provider for sufficiently-popular-but-still-niche models.
The main reason to use featherless is to experiment and use a wider variety of fine-tunes than anywhere else.
@taishiling - im glad you like being able to play with all the models.
Yes, we are currently downloading more models which would be coming online over the next few days.
We currently support RWKV and LLaMA based models. We plan to introduce Mistral MoE's next (to be confirmed), followed by potentially larger models.
The main reason we started with llama, is because it was the largest pool of all the popular models our initial users and community wanted to use.
But the goal remains: ALL huggingface models. One major group at a time
Hey Team Featherless.AI,
Congratulations on the launch of Featherless.AI! 🎉
This is a game-changer for anyone looking to access top models on Hugging Face without the hassle and high costs of setting up and renting GPUs. By offering a wide range of models with simple and affordable pricing plans, you’re making cutting-edge AI technology accessible to everyone.
The focus on data security and privacy is commendable, and the custom-built infrastructure to dynamically hot-swap models is truly innovative. This ensures users can explore a vast array of models seamlessly.
Excited to see how Featherless.AI transforms the AI landscape!
How does Featherless ensure the integration of these models remains user-friendly?
It sounds like an amazing platform! The ability to access the latest open-source AI models from Hugging Face is also fantastic.
@hakz Thanks!
We use OpenAI compatible API, so any tools that integrates and work with OpenAI can be used with featherless.ai
Congrats on the launch! Featherless AI offers great access to Hugging Face models at a fair price Excited to see what it can do!
@andreas_sohns Thanks!
Next step: More models!
After that: To make it easy for anyone to finetune their own models, for their use case. And use it =)
Hi Eugene, Featherless.AI sounds like an invaluable tool for accessing and experimenting with various LLMs. How do you manage the computational resources to support such a wide range of models? Also, are there plans to support custom model training or fine-tuning on the platform? Congrats!
@kyrylosilin
Our infrastructure auto scale according to current workload, with optimizations specific to our infra provider.
We also adjust our setup for the best price to performance GPUs, to lower overall cost (aka no H100s), even for the larger models.
In overall, this lets us keep costs to the minimum (no wasted GPUs), according to the user workloads.
And yes, we do plan to support additional features like finetuning, and the option to switch to token based pricing, for larger scale commercial users in the future. But for now, this is aimed at individuals who would love to play with all the models
Inoticed your pricing details did not mention anything about API access. Perhaps you could update it so non PH users can see it too :)
Congrats on the launch!
Great product, very useful to be able to use and change llama models instantly, without having to set up and deploy a server myself!!
Congrats on the launch: I was wondering how this is feasible with the cost of gpus and hosting this many models?
@nathan_wilce Thanks!
All the models are pre-downloaded to our clusters, and are on standby.
The GPUs spin's up and swap these models on the fly under seconds.
This was made possible with our custom built inference systems we built and optimized on.
Allowing us to keep cost low, and auto scale to actual usage.
PS: 450+ models takes up over 9TB of storage space, we are downloading 100+ more which will be coming online soon. (it takes time haha)
@julien_c Thank you 🙏 : now we just need to scale usage, then servers, then models - until we catch them all from huggingface 😉
@julien_c
Thanks! They are actually the same backend / infra 😉
As we scale featherless.ai, we do plan to integrate it back to recursal, with the more "complicated features" like automated finetuning, etc. So that recursal will be providing offerings beyond RWKV.
@tim__pouw Thanks, hope you can find a useful model for your use case, or for fun!
Featherless AI