Scaleway added as Hugging Face Inference Provider – capabilities, usage, and billing
TL;DR
Scaleway is now an official Inference Provider on the Hugging Face Hub, letting users run frontier models such as GPT‑OSS, Qwen3, DeepSeek R1 and Gemma 3 via server‑less APIs with European‑hosted, low‑latency infrastructure.
Why the addition matters
Adding Scaleway expands the geographic and pricing options for Hugging Face inference, offering sub‑200 ms first‑token latency, pay‑per‑token rates starting at €0.20 M tokens, and direct routing of requests through a European data center. This gives European developers data‑sovereignty guarantees and more flexibility in choosing between Hugging Face‑routed billing and direct provider billing.
Supported models and access points
- Popular open‑weight models are immediately available with the
inference_provider=scalewayquery parameter, e.g.openai/gpt-oss-120b,Qwen/Qwen3-Coder-30B-A3B-Instruct,deepseek-ai/DeepSeek-R1-Distill-Llama-70B, andgoogle/gemma-3-27b-it. - The Scaleway organization page on the Hub (
https://huggingface.co/scaleway) lists all models hosted by the provider. - Trending Scaleway‑backed models can be browsed at
https://huggingface.co/models?inference_provider=scaleway&sort=trending.
How it works
In the Hub UI
- API key management – In user settings you can store a custom Scaleway API key. If no key is set, requests are automatically routed through Hugging Face, which bills your HF account.
- Provider preference – You can order providers by preference; the model widget and code snippets respect this ordering.
- Model page widget – Each model page shows compatible third‑party providers, sorted according to your preferences.
From client SDKs
Python (huggingface_hub ≥ 0.34.6)
import os
from huggingface_hub import InferenceClient
client = InferenceClient(
provider="scaleway",
api_key=os.environ["HF_TOKEN"],
)
messages = [{"role": "user", "content": "Write a poem in the style of Shakespeare"}]
completion = client.chat.completions.create(
model="openai/gpt-oss-120b",
messages=messages,
)
print(completion.choices[0].message)
You may replace api_key with a raw Scaleway key to bypass HF routing.
JavaScript (@huggingface/inference)
import { InferenceClient } from "@huggingface/inference";
const client = new InferenceClient(process.env.HF_TOKEN);
const chatCompletion = await client.chatCompletion({
model: "openai/gpt-oss-120b",
messages: [{ role: "user", content: "Write a poem in the style of Shakespeare" }],
provider: "scaleway",
});
console.log(chatCompletion.choices[0].message);
Both snippets demonstrate the same API surface; the only difference is the provider field.
Billing model
- Direct requests – When you supply a Scaleway API key, billing is handled by Scaleway on your account.
- Routed requests – When you authenticate with a Hugging Face token, the Hub forwards the request and charges you the provider’s standard API rate with no additional markup.
- PRO credits – Hugging Face PRO users receive $2 of inference credits each month, usable across any provider, including Scaleway.
- Free tier – Signed‑in free users get a limited free quota; upgrading to PRO unlocks higher limits and ZeroGPU access.
Implications for developers
- Data sovereignty – Hosting inference in Paris‑based data centers satisfies EU‑centric compliance requirements.
- Performance – Sub‑200 ms first‑token latency makes Scaleway suitable for interactive chatbots and agentic pipelines.
- Cost flexibility – Pay‑per‑token pricing and the ability to choose between direct and routed billing let teams optimise spend.
- Multimodal support – The service handles structured outputs, function calling, and both text and image generation, expanding use‑case coverage.
Next steps and feedback
- Review the dedicated documentation at
https://huggingface.co/docs/inference-providers/providers/scalewayfor advanced configuration. - Test the trending model list to evaluate latency and cost for your workloads.
- Provide feedback via the Hugging Face discussion space:
https://huggingface.co/spaces/huggingface/HuggingDiscussions/discussions/49.
This post summarizes the official Hugging Face announcement dated 2025‑09‑19.