Baseten in the Hugging Face ecosystem: what it means
The Baseten platform has officially become one of the inference providers on the Hugging Face Hub. This is an important step for everyone working with open models: now you can launch them directly from the model page without deploying your own infrastructure. Previously, you had to go to third-party services, copy keys, and manually configure the environment — now the process is reduced to a few clicks.
Baseten is not just hosting for LLMs. The platform offers serverless inference, model training, and a catalog of ready-made neural networks — from large language models to speech synthesis systems. At the initial stage, the Hugging Face integration covers tasks for conversational assistants and text generation. Available open-weight models include Kimi K3, DeepSeek V4 Flash, GLM-5.2, and others. The list of supported scenarios is expected to expand later.
How to set up model deployment
Users have two ways to access the provider. The first is to specify your own Baseten API key in your account settings on Hugging Face. Then requests from model pages go directly to the service, and you pay for them according to your plan. The second mode is routing through Hugging Face infrastructure. In this case, a separate Baseten key is not needed: authentication happens via your Hugging Face token, and all charges are billed to your platform account.
Choosing a provider is easy to control: you can set the order of preferences in your personal settings. If a model is available from several third-party services, Hugging Face will automatically show them on the model page, sorted according to your settings. This is convenient when you want to compare the speed or cost of different infrastructures.
Client libraries and agent frameworks
Developers won't have to write additional boilerplate. Baseten is already available through official SDKs: the huggingface_hub library version 1.26.1 and above for Python, as well as the @huggingface/inference package for JavaScript. Just authenticate with your Hugging Face token — and requests to models hosted on Baseten will be routed automatically.
There's also good news for those building agent systems. Inference providers are integrated with popular agent harnesses, including Pi, OpenCode, Hermes Agents, and OpenClaw. This means a model on Baseten can be used in such tools without writing adapter modules — the connection happens through the framework's standard means.
Cost and free limits
The question of money is resolved differently depending on the chosen mode. If you use your own provider key, billing is entirely on the Baseten side. When routing through Hugging Face, the provider's standard rates apply without any markups from the platform. Revenue-sharing agreements may be possible in the future, but for now it's a standard pay-as-you-go scheme.
Active users get nice bonuses. Hugging Face PRO subscribers receive $2 per month in inference credits — they work across different providers. And registered free users get a small amount of inference at no charge, so you can try the new integration without any initial investment.
The appearance of Baseten in the list of Inference Providers makes the Hugging Face ecosystem even more flexible. Users get more choice, and developers get less routine. To deploy open LLMs, you no longer need to be a DevOps expert: a couple of settings are enough, and the model is up and running.
