Hugging Face Adds Baseten as Serverless Inference Provider

Hugging Face has expanded its serverless inference ecosystem by adding Baseten as a supported third-party provider. This integration allows developers to run popular open-weight large language models directly from model pages on the Hugging Face Hub. At launch, the partnership focuses on conversational and text-generation tasks, supporting frontier models such as GLM-5.2, Kimi K3, and DeepSeek V4 Flash.

Developers can access Baseten's infrastructure through Hugging Face's client SDKs, specifically requiring version 1.26.1 or higher of the Python huggingface_hub library, or the @huggingface/inference package for JavaScript. The setup supports popular agent frameworks including Pi, OpenCode, Hermes Agents, and OpenClaw. This allows practitioners to deploy Baseten-hosted models into their existing workflows without writing custom integration code.

The system offers two distinct routing and billing pathways. Users can configure their own Baseten API keys to route calls directly to the provider, in which case billing is handled through their Baseten account. Alternatively, requests can be routed through Hugging Face using the base URL router.huggingface.co/v1. For routed requests, Hugging Face passes through Baseten's standard pricing rates with no additional markup, charging the user's Hugging Face account directly.

To encourage adoption, Hugging Face is offering its Pro subscribers two dollars in monthly inference credits, which can be applied toward Baseten or other third-party providers. Free tier users also receive a small daily quota for testing. Users can manage their preferred providers and API keys directly within their Hugging Face account settings, which automatically updates the code snippets and widgets displayed on model pages.