Cohere now serves its own enterprise models directly on Hugging Face Hub

Hugging Face just landed a first for its Inference Providers program: Cohere is not only joining the roster, but it’s the first model creator to serve its own models directly on the Hub. This isn’t just another API integration. It means the team that built Command and Aya is now the one running them for you, through a serverless endpoint. Starting today, nine models are available, spanning the high-powered enterprise workhorses and the multilingual research lab experiments.

You can now fire off calls to the new CohereLabs/c4ai-command-a-03-2025, a model built for demanding business applications with a massive 256k context window—double what most leading models offer. It’s designed for verifiable RAG with citations and agentic tool use. If you need something leaner, the c4ai-command-r7b-12-2024 targets low-cost, low-latency scenarios without ditching the 128k context or 23-language support. Cohere’s not leaving its research side behind either. The Aya Expanse and Aya Vision families are live, headlined by the 32-billion parameter Aya Vision model that extends multimodal capabilities across 23 languages.

Getting started is straightforward from Python. You’ll need huggingface_hub v0.30.0 or later, and you can route through a standard Hugging Face token or bring your own Cohere API key. The workflow in the SDK is simple: define “cohere” as your provider when initializing InferenceClient, then call the chat completion endpoint with the model ID. Vision tasks work the same way—encode your image as a base64 data URL and drop it into the messages array alongside text. You can also browse and test models directly in the Hub UI by filtering for the Cohere provider, which gives you a playground without touching code.

What’s interesting here is the signal this sends about distribution. Model builders typically rely on platforms like Hugging Face to host community weights, but serving inference is often kept in-house or outsourced to a major cloud. Cohere planting its flag as a provider blurs that line. For developers, it means you’re getting the model served by the people who understand its quirks best, not a generic third-party wrapper. The real test will be latency and reliability at scale, but the setup at least removes a layer of abstraction between the model creator and the end user.