Fireworks.ai lands on Hugging Face Hub with serverless access to DeepSeek-R1

Hugging Face just gave developers a serious speed boost by integrating Fireworks.ai directly into its Hub. Starting now, you can run serverless inference on a curated list of models — including DeepSeek-R1, DeepSeek-V3, Mistral-Small-24B, Qwen2.5-Coder-32B, and Llama-3.2-90B-Vision — without ever leaving the Hugging Face ecosystem. The integration works across the website UI, Python and JavaScript client libraries, and raw HTTP calls, which means it slots into existing workflows with minimal friction.

The mechanism is straightforward: when using huggingface_hub in Python or @huggingface/inference in JavaScript, you simply specify provider="fireworks-ai" and the request routes through Fireworks’ infrastructure. You can authenticate with either a Hugging Face token for automatic routing or bring your own Fireworks API key for direct billing. The cURL examples show the same pattern — just point at router.huggingface.co/fireworks-ai and pass your credentials. For devs already deep in the HF ecosystem, this eliminates the step of managing a separate inference endpoint.

Billing is refreshingly transparent. Direct requests using a Fireworks key land on your Fireworks account. Routed requests through the Hub pass through Fireworks’ standard API rates with no Hugging Face markup — at least for now. The post hints at future revenue-sharing agreements with provider partners, which could change the economics down the line, but today it’s a clean passthrough. PRO subscribers also get $2 in monthly inference credits usable across providers, which isn’t life-changing money but covers a decent chunk of experimentation.

What makes this notable isn’t just another inference option — it’s the models Fireworks brought along. DeepSeek-R1 has become a go-to for reasoning-heavy tasks, and having it available serverless on the Hub lowers the barrier for prototyping with a model that typically demands real compute. The Mistral-Small-24B and Qwen2.5-Coder-32B inclusions suggest Fireworks is targeting developers who need fast, cost-effective inference for coding and general-purpose tasks. Given how crowded the inference provider space has become — with Together, Replicate, and Groq all competing on speed — Fireworks is betting that tight Hugging Face integration, not just raw throughput, will win over developers who value workflow simplicity over benchmark bragging rights.