Love that I can call DeepSeek R1 directly from the Hub 🔥

from huggingface_hub import InferenceClient

client = InferenceClient(
    provider="together",
    api_key="xxxxxxxxxxxxxxxxxxxxxxxx"
)

messages = [
    {
        "role": "user",
        "content": "What is the capital of France?"
    }
]

completion = client.chat.completions.create(
    model="deepseek-ai/DeepSeek-R1", 
    messages=messages, 
    max_tokens=500
)

print(completion.choices[0].message)

Is it possible to set a monthly payment budget or rate limits for all the external providers? I don't see such options in billings tab. In case a key is or session token is stolen, it can be quite dangerous to my thin wallet:(

@benhaotang you already get spending notifications when crossing important thresholds ($10, $100, $1,000) but we'll add spending limits in the future

@benhaotang you already get spending notifications when crossing important thresholds ($10, $100, $1,000) but we'll add spending limits in the future

Thanks for your quick reply, good to know!

Would be great if you could add Nebius AI Studio to the list :) New inference provider on the market, with the absolute cheapest prices and the highest rate limits...

Could be good to add featherless.ai

TitanML !!

Hey @julien-c ,
I am from Chutes.ai, a completely open source focused inference provider (even our own code is open).
We'd love to join HF as a inference provider if possible.
Any chance to get onboarded?
You find me here or via florian@chutes.ai or via X DMs at https://x.com/airesearch12.
Thanks
Florian

I am trying to perform an experiment for a research project, and I need access to nvidia/Llama-3.1-Nemotron-70B-Instruct-HF through an inference provider. I do not have the resources to run the model locally. About a week ago, nvidia/Llama-3.1-Nemotron-70B-Instruct-HF was available through the provider featherless-ai. Now, no providers are serving the model. Please help. thanks.

We evaluate provider-routed calls using sanitized, non-sensitive test prompts. Is there a supported per-call field, header, or API that supplies an opaque correlation handle, the exact immutable served-model revision or artifact digest, and a verifiable binding between them (for example, a signed attestation)? Static Hub metadata is insufficient for routed inference. If unsupported, please confirm and identify the feature-request path.

Inference Provider Request: Aster (Kimi K3 + GLM 5.2 + More to come...)

Hi HF team 👋

We'd love to list Aster as an inference provider; we serve open-weight models on our infra.

We currently serve:

• zai-org/GLM-5.2
• moonshotai/Kimi-K3 : Image inputs and conversation endpoints work on our endpoint (image-text-to-text / conversational).
• openai/gpt-oss-120b

We've followed the provider docs and:

• Are registered on the Team plan @ aster-ai.
• Have an OpenAI-compatible API (/v1/chat/completions, and /v1/models already serving pricing + ctx length)
• We've implemented the Inference-Id response header & billing callback

Ready to open the PRs once our organization is enabled.

May you let us know how to enable our organization on the Model Mapping API, and who to contact for the payout setup?

API base URL: https://api.asterlab.ai
Contact: info@asterlab.ai

Thanks 🤗

Hi HF team and @julien-c

We'd love to be added as an Inference Provider: Infersia.

We serve open-weight models on dedicated GPUs through an OpenAI-compatible endpoint (https://api.infersia.com/v1), and we've worked through the register-as-a-provider guide:

Provider id: infersia · Org (Team plan): https://huggingface.co/infersia-com
JS client PR: https://github.com/huggingface/huggingface.js/pull/2343 — BaseConversationalTask, no payload overrides needed; tsc/eslint/oxfmt clean, and both new tests verified against the live endpoint
Inference-Id header on every response, including streaming ones
Billing endpoint implemented and returning integer nano-USD — that is our ledger's native unit, so the figure is exact rather than converted
/v1/models publishes pricing.input / pricing.output in USD per million, plus context_length, for the provider comparison table
Two things we think make us worth a slot rather than just another endpoint:

We serve full context windows. DeepSeek V4 Flash runs at its complete 1,048,576 tokens rather than a truncated slice.

We publish the quantisation. Every endpoint states the precision it is served at, and the hardware behind it. It is the biggest hidden variable in commercial inference and almost nobody discloses it, which is why the same model behaves differently across providers and people blame their own code.

Current catalogue: DeepSeek V4 Flash, Qwen3.6 35B A3B, Qwen3 14B, Qwen3 8B, StepFun Step 3.7 Flash (vision). There is also a free tier — qwen/qwen3-8b:free, no card. More models on the way, but serving these with low ttft and high tps right now.

Could someone enable us server-side so we can start registering model mappings in staging? That is the one step we can't do ourselves.

Happy to provide anything else that would help. Thanks!

— Scott, Infersia

PS: @Wauplin Lucain and @celinah Célina please let us know directly what we can do to make this work from you end. We've got compute and are flexible.

Any inference provider you love, and that you'd like to be able to access directly from the Hub?

Hi! I would be very happy if you would add the NEURALWATT.COM
It is very very cheap
Thanks in advance!!!

Hi @julien-c and the Hugging Face team,

We would like to apply to become a Hugging Face Inference Provider for automatic speech recognition.

Proposed provider ID: vibevoice-asr
Hugging Face organization: https://huggingface.co/vibevoice-asr
Website: https://vibevoice-asr.com
Initial model: microsoft/VibeVoice-ASR
Task: automatic-speech-recognition
Contact: hello@mail.vibevoice-asr.com

We are building a dedicated inference service for VibeVoice-ASR and intend to support Hugging Face Routed Requests.

Implementation is already underway, and we expect to have a staging-ready endpoint, initial pricing, and preliminary performance results available during the week of August 10, 2026.

We are following the official provider registration guide and plan to provide:

• Hugging Face ASR task-compatible input and output handling
• an Inference-Id header for every successful response
• a billing endpoint returning costNanoUsd
• a huggingface.js provider integration, followed by huggingface_hub integration once the Hub integration is operational
• a staging model mapping backed by a publicly reachable endpoint for Hugging Face automated validation
• usage-based pricing calculated from audio duration

We would like to begin the onboarding process early so that we can align our implementation with Hugging Face requirements before our staging endpoint is ready.

We understand that a Team or Enterprise organization plan is required before accessing the Model Mapping API, and we are prepared to upgrade before reaching that stage.

Could you please advise:

1. whether our organization can begin the provider onboarding process while the infrastructure is being completed;
2. who we should coordinate with for server-side enablement, commercial arrangements, and payout onboarding;
3. whether there are any ASR-specific requirements beyond the published provider guide.

We will share our staging endpoint, pricing, performance results, and integration PRs as soon as they are ready.

We are an independent inference provider and are not affiliated with or endorsed by Microsoft.

Thank you!