Inference APIs
Reference/Errors/Hugging Face

"Model … inference is not supported HF inference api" / "not available via any of the supported Inference Providers" (Hugging Face 404)

The model page loads, the id is right, the token is fine, and the API still answers 404 Model … inference is not supported HF inference api or This model is not currently available via any of the supported Inference Providers. This is the difference between the old serverless Inference API, which would load almost any public checkpoint on demand, and today's Inference Providers, which only serve models that a provider has deployed. Nothing about your request is wrong; the model is not hosted.

Last verified November 20, 2026 against HF forum: Inference API stopped working (Apr 2025), HF forum: not available via any supported Inference Providers, HF Inference Providers docs · 4 min read

ProviderHugging Face
HTTP status404 Not Found (sometimes 400)
MessagesModel <id> inference is not supported HF inference api · This model is not currently available via any of the supported Inference Providers · Model <id> does not exist
When it happensThe model exists on the Hub but no Inference Provider (including hf-inference) has it deployed
Can you wait it out?Not usefully. Models are added when a provider chooses to host them, with no schedule
Short answer
  • Check the model page. The Inference Providers widget on the Hub lists which hosts serve it; if the widget is missing, no one does.
  • Pin a provider with provider="…" or a :provider suffix when one host serves it and routing picked another.
  • Standard open model? Run the same request against a host that serves it at a fixed price.

Why this happens

Until 2025 api-inference.huggingface.co would spin up most public models on request (you saw Model is loading and a 503 for a minute, then answers). The replacement, Inference Providers, forwards each request to a third-party host, and a host serves only the models it has chosen to deploy. A fine-tune with 40 downloads, an older base model or a research checkpoint usually has no provider, and the 404 above is the answer. The hf-inference provider (Hugging Face's own) likewise serves only a curated, warm set; its client library warns that it will only support warm models in the future (no cold start allowed).

Option 1 — find a model that is served

Open the model page on the Hub. A served model shows an Inference Providers panel with the hosts and a code snippet. Programmatically:

Python
from huggingface_hub import HfApi

api = HfApi()
for m in api.list_models(inference_provider="together", task="text-generation", limit=20):
    print(m.id)
# any provider:  api.list_models(inference_provider="all", ...)

If your model is a small fine-tune, the pragmatic move is the base model it was tuned from, which is far more likely to be deployed.

Option 2 — pin the provider

When one provider serves the model and the router picked another (or the default hf-inference), say which one.

Python
from huggingface_hub import InferenceClient

client = InferenceClient(provider="together", api_key=os.environ["HF_TOKEN"])
out = client.chat_completion(model="meta-llama/Llama-3.3-70B-Instruct", messages=[{"role": "user", "content": "Hello"}])

# OpenAI SDK against the router: suffix the model id
# model="meta-llama/Llama-3.3-70B-Instruct:together"

Option 3 — run the standard open model elsewhere

If the model you wanted is one of the widely hosted open-weight models, you do not need Hugging Face in the path at all: the request is the OpenAI chat call, so it runs unchanged against any host that serves the model. Inference APIs publishes a price per model, bills per request from a prepaid balance, and accepts the Hub-style ids.

Python
from openai import OpenAI

client = OpenAI(base_url="https://api.inferenceapis.com/v1", api_key=os.environ["INFERENCEAPIS_API_KEY"])
resp = client.chat.completions.create(
    model="Qwen/Qwen3.8-Flash",   # $0.15 / $0.50 per 1M tokens; also Llama 3.3 70B, DeepSeek V4, GPT-OSS, GLM, Kimi
    messages=[{"role": "user", "content": "Hello"}],
)

Whisper Large v3 (openai/whisper-large-v3) and Kokoro (hexgrad/Kokoro-82M), two Hub checkpoints people often called on the old host, are served the same way under /v1/audio/transcriptions and /v1/audio/speech.

Frequently asked questions

Can I ask Hugging Face to host the model?

Yes: the model page has a request link for provider support, and HF staff point people to it. There is no timeline, so plan as if it will not happen.

Does a PRO subscription make more models available?

No. PRO adds monthly credit and higher quotas, not models. Availability depends on providers.

Why did this model work last month?

Providers drop models. The hf-inference provider in particular is moving to warm models only, and a model that was served cold can stop being served without notice.

Where Inference APIs fits

Option 3 is for the common case where the model is a standard open-weight checkpoint: Inference APIs serves Llama 3.3 70B, Qwen3, DeepSeek V4, GPT-OSS, GLM, Kimi, Whisper Large v3 and Kokoro behind the same OpenAI-style call, at a fixed price per model.

Something changed or wrong? Tell us and we will re-verify the entry.