Baseten Joins Hugging Face Inference Providers for Serverless Open-Weight LLM Access

Summary

Hugging Face announced on August 6, 2026, that Baseten had joined its Inference Providers program, bringing Baseten-hosted models into the Hugging Face Hub. The integration lets developers invoke compatible models from Hub model pages, Hugging Face client libraries, and supported agent tools. Baseten is an AI infrastructure company offering serverless inference, model training, and a catalog spanning multiple AI modalities. At launch, however, the Hugging Face integration is limited to conversational and text-generation workloads. The initial catalog includes open-weight large language models identified by the post as Kimi K3, DeepSeek V4 Flash, and GLM-5.2, alongside other models available through Baseten. Hugging Face demonstrates the integration with the model identifier “deepseek-ai/DeepSeek-V4-Flash-0731:baseten.” The supplied article links to a complete catalog but does not enumerate it, so the exact number of supported models cannot be established from the text alone. Hugging Face says more task types will be introduced, although it gives no schedule or list of which capabilities will come next. Users can access Baseten in two ways. With the custom-key option, a developer supplies a Baseten API key, the request goes directly to Baseten, and charges appear on the Baseten account. With Hugging Face routing, the developer authenticates using a Hugging Face token, does not need separate Baseten credentials, and receives the provider charge through the Hugging Face account. Users can also rank providers by preference, which influences provider selection in model-page widgets and generated code examples. Programmatic access is available through the Python package huggingface_hub version 1.26.1 or later and the JavaScript package @huggingface/inference. The article also shows OpenAI-compatible Python and JavaScript requests sent through Hugging Face’s router endpoint. Hugging Face says its provider framework works with agent harnesses including Pi, OpenCode, Hermes Agents, and OpenClaw, allowing Baseten-hosted models to be connected without bespoke integration code. The addition gives developers another deployment option while making it easier to compare or switch infrastructure providers behind a common Hugging Face interface. That could reduce integration work for application teams already using the Hub, especially when they want centralized authentication or billing. Hugging Face says routed usage carries the provider’s standard API price without an additional markup, while PRO subscribers receive $2 in monthly inference credits and signed-in free users receive an unspecified small quota. The post does not disclose model-specific prices, latency, regional availability, capacity, service-level guarantees, or performance benchmarks, so the practical competitiveness of Baseten cannot be judged from the announcement alone.

What happens next remains partly undefined. Hugging Face is soliciting feedback through its discussion forum, and Baseten is expected to add support beyond chat and text generation. There is no published rollout date, and Hugging Face notes that it may eventually establish revenue-sharing arrangements with providers. The post does not say whether that possibility would affect customer pricing. Consequently, the integration is immediately useful for supported models, but its broader value will depend on catalog growth, production reliability, and future commercial terms.