Inference providers seems to not be compatible with `text_generation`

[anakin87]

Describe the bug

I started from this table, where it seems that some inference providers are compatible with text_generation.

Then, taking nebius as an example, I filtered text-generation models served by Nebius.

But when I try to call the client, I get
ValueError: Model Qwen/Qwen3-4B is not supported for task text-generation and provider nebius. Supported task: conversational.

This also applies to other models and Inference providers.

Reproduction

import os
from huggingface_hub import InferenceClient

client = InferenceClient(
    model="Qwen/Qwen3-4B",
    provider="nebius",
    api_key=os.environ["HF_TOKEN"],
)

text = client.text_generation(
    prompt="What is the capital of France?",
)

print(text)

Logs

Send: curl -X GET -H 'Accept: */*' -H 'Accept-Encoding: gzip, deflate' -H 'Connection: keep-alive' -H 'authorization: <TOKEN>' -H 'user-agent: unknown/None; hf_hub/0.33.2; python/3.12.8; torch/2.6.0' 'https://huggingface.co/api/models/Qwen/Qwen3-4B?expand=inferenceProviderMapping'
Request 52984ed5-11a7-4a01-a565-d0f14d6d01c0: GET https://huggingface.co/api/models/Qwen/Qwen3-4B?expand=inferenceProviderMapping (authenticated: True)
Traceback (most recent call last):
  File ".../try.py", line 10, in <module>
    text = client.text_generation(
           ^^^^^^^^^^^^^^^^^^^^^^^
  File ".../test/lib/python3.12/site-packages/huggingface_hub/inference/_client.py", line 2297, in text_generation
    request_parameters = provider_helper.prepare_request(
                         ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File ".../test/lib/python3.12/site-packages/huggingface_hub/inference/_providers/_common.py", line 93, in prepare_request
    provider_mapping_info = self._prepare_mapping_info(model)
                            ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File ".../test/lib/python3.12/site-packages/huggingface_hub/inference/_providers/_common.py", line 162, in _prepare_mapping_info
    raise ValueError(
ValueError: Model Qwen/Qwen3-4B is not supported for task text-generation and provider nebius. Supported task: conversational.

System info

- huggingface_hub version: 0.33.2
- Platform: macOS-15.5-arm64-arm-64bit
- Python version: 3.12.8
- Running in iPython ?: No
- Running in notebook ?: No
- Running in Google Colab ?: No
- Running in Google Colab Enterprise ?: No
- Token path ?: ...
- Has saved token ?: True
- Who am I ?: anakin87
- Configured git credential helpers: osxkeychain, store
- FastAI: N/A
- Tensorflow: N/A
- Torch: 2.6.0
- Jinja2: 3.1.6
- Graphviz: N/A
- keras: N/A
- Pydot: N/A
- Pillow: 11.3.0
- hf_transfer: N/A
- gradio: N/A
- tensorboard: N/A
- numpy: 2.2.6
- pydantic: 2.11.7
- aiohttp: 3.12.13
- hf_xet: 1.1.5
- ENDPOINT: https://huggingface.co
- HF_HUB_CACHE: ...
- HF_ASSETS_CACHE: ...
- HF_TOKEN_PATH: ...
- HF_STORED_TOKENS_PATH: ...
- HF_HUB_OFFLINE: False
- HF_HUB_DISABLE_TELEMETRY: False
- HF_HUB_DISABLE_PROGRESS_BARS: None
- HF_HUB_DISABLE_SYMLINKS_WARNING: False
- HF_HUB_DISABLE_EXPERIMENTAL_WARNING: False
- HF_HUB_DISABLE_IMPLICIT_TOKEN: False
- HF_HUB_ENABLE_HF_TRANSFER: False
- HF_HUB_ETAG_TIMEOUT: 10
- HF_HUB_DOWNLOAD_TIMEOUT: 10

[hanouticelina]

hey @anakin87,
thanks for the detailed report and very sorry about this.
What's happening here is that Qwen/Qwen3-4B has a tag "conversational" - which means that even if the model task is text generation, the model is only served as a chat completion model, i.e.
(task =text-generation) + (model tagged as conversational) = served as chat completion. The only exception is with the HF Inference API provider, where we allow both text generation and chat completion calls.

That's why calling text_generation() fails here, Nebius does not serve that model for the text-generation task, only for conversational. The following should work as expected:

import os
from huggingface_hub import InferenceClient

client = InferenceClient(
    provider="nebius",
    api_key=os.environ["HF_TOKEN"],
)

completion = client.chat.completions.create(
    model="Qwen/Qwen3-4B",
    messages=[
        {
            "role": "user",
            "content": "What is the capital of France?"
        }
    ],
)

print(completion.choices[0].message)

Besides the "conversational" tag, you can refer to the inference snippets on the model page to double-check how the model is served and which task method to use.

Side Note: today most text generation models on the Hub are chat models.

I'm closing this issue, but feel free to comment if you have any follow up questions or suggestions on how we can improve things! 🤗

[anakin87]

Ok, thanks for explanation.

I would maybe suggest updating this table: https://huggingface.co/docs/huggingface_hub/guides/inference#supported-providers-and-tasks

[hanouticelina]

ah yes forgot to mention: the table is actually correct, it only shows which tasks are supported by which providers, and currently mistralai/Devstral-Small-2505 is the only model supported for text generation task (not conversational) by Nebius.
sorry for the confusion!