Spaces:

---

huggingface

/

HuggingDiscussions

Like 104

Running

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 !!

MuninCloud (Bogarra Integrated SL, Spain): request to join Inference Providers

Hi Hugging Face team,

We are MuninCloud (https://munincloud.com), the cloud brand of Bogarra Integrated SL, a Spanish company, and we would like to join Inference Providers.

What is live today

• OpenAI-compatible API: https://api-flex-huggingface.munincloud.com/v1 (chat completions, streaming with token usage in streaming and non-streaming responses).
• /v1/models with pricing (USD per million tokens) and context_length: https://api-flex-huggingface.munincloud.com/v1/models
• Model: Qwen/Qwen3-32B served as AWQ 4-bit with vLLM, context window 16,384 tokens, up to 4,096 output tokens. Price: $0.11 / $0.39 per million input / output tokens.
• Dedicated servers in Germany (Hetzner, Falkenstein), inside the EU.
• Zero data retention: prompts and completions are never written to disk or logs and never used for training. Prefix cache is salted per account (measured: no cross-account cache hits).
• An Inference-Id header (UUID) on every response.
• When our engine is busy we answer 429 with Retry-After, never a 5xx.
• Privacy policy: https://munincloud.com/legal/flex/en/privacy · Terms: https://munincloud.com/legal/flex/en/terms

What we will do next, following the provider guide

• Open the huggingface.js PR (OpenAI-compatible conversational helper).
• Implement the billing endpoint (cost per request ID in nano-USD).
• Upgrade our Hub organization AdminMunin (https://huggingface.co/AdminMunin) to the Team plan when the mapping step is enabled.

Questions

1. How and in which currency is routed usage settled with the provider?
2. Is an AWQ 4-bit deployment acceptable mapped to Qwen/Qwen3-32B (we state the quantization openly)?
3. Does your router read /v1/models with our provider key, or without authentication?
4. What are the next steps to enable our provider account server-side?

Contact: javier.bogarra@munincloud.com

Thank you,
The MuninCloud team

Hi Hugging Face team — I’m Gaurav Sharma, founder of SaaS Labs, and we’d like to bring PyAI Hear (automatic speech recognition) and Speak (text-to-speech) to Inference Providers. Hub organization: https://huggingface.co/pyaibyatomsai.

PyAI helps developers add speech input and spoken answers to their existing apps and agents while keeping their chosen LLM and orchestration. Hear supports 29 recognition languages with automatic detection. Speak’s live catalog offers 200 stock voices across 17 language groups. These are supported-language/catalog counts, not a claim of equal measured quality in every language. Our LiveKit and Pipecat integrations are already published on PyPI.

Current direct pricing: Hear $0.001 per audio minute (synchronous nonempty requests round up to started minutes; empty transcripts are free). Speak $0.04 per generated audio minute, using actual output duration. We’d like to agree audio-specific billing and settlement with your team.

A recent bounded production check generated a 2.48-second WAV in 1,294 ms and transcribed the exact sentence in 1,530 ms (JSON) / 2,439 ms (plain text). This is one short English example and full-response timing—not first-audio latency, a percentile benchmark or SLA. PyAI’s default no-training warranty covers customer inputs and outputs; it is separate from retention, and we do not claim zero retention.

API: https://api.pyai.com/v1 (authenticated OpenAI-compatible audio endpoints). Docs: https://docs.pyai.com. Pricing: https://pyai.com/pricing. Privacy: https://pyai.com/legal/privacy.

Could you confirm eligibility and the model-mapping path for these API products before we submit the JS/Python adapters? In particular: are API-only speech products eligible, which Hub model repositories should they map to, and what are the billing precision and organization-plan requirements for qualification? We can then implement the agreed task schemas and billing endpoint. This is an onboarding request; PyAI is not yet a native HF Inference Provider.

Contact: gaurav@pyai.com. Thank you!

Following up on PyAI’s onboarding request with new, reproducible integration evidence. Our goal is to bring Hear (ASR) and Speak (TTS) to Inference Providers, so developers can add speech input and spoken answers around their chosen LLM.

• Live AI SDK 7 guide
• Public runnable source and six passing offline contract tests
• Passing public CI

The exact AI SDK example passed a production Speak → Hear round trip: HTTP 200 on both requests, no retries, a valid 2.48-second WAV (119,084 bytes), and the correct transcript “Hello, your voice integration is ready.” Full-response timings were 1,954 ms for Speak and 1,464 ms for Hear. This is one short English functional sample—not first-audio latency, a percentile benchmark or an SLA. It verifies the direct PyAI integration, not native Hugging Face routing.

Current direct rates remain Hear $0.001 per started input-audio minute per nonempty synchronous request (empty transcripts free) and Speak $0.04 per generated minute, using actual output duration. Hear supports 29 recognition languages; Speak’s catalog contains 200 stock voices across 17 language groups. These are support/catalog counts, not equal-quality claims. Our default no-training warranty covers inputs and outputs, separately from retention.

Could the provider onboarding team confirm whether these API-only speech products are eligible and which Hub model repositories they should map to? We’ve reviewed the provider guide and understand the remaining JS task adapter, request-cost billing endpoint, server-side mappings and organization-plan steps. We want to implement the agreed ASR/TTS contract rather than assume the LLM compatibility path applies.

An onboarding owner and the next qualification checklist would help us move directly to a focused pilot. Organization: pyaibyatomsai. Contact: gaurav@pyai.com. Thank you!

Quick update to our note above (Oct 18) - two things changed since then:

• We added a vision-capable model to the lineup: Qwen/Qwen3-VL-30B-A3B-Instruct ($0.14 / $0.57 per 1M input/output tokens, 32K context, OpenAI-compatible chat completions).
• We published a Space here on the Hub so people can try our models without leaving it: https://huggingface.co/spaces/linkrra-si/synap

Live catalog + pricing (OpenAI-compatible), always current: https://synap.linkrra.com/v1/models

Still happy to open the huggingface.js PR and follow the registration guide whenever useful on your side - no rush.

Thanks!
The Linkrra team

Following up on our GoedelMachines provider request: we have moved from Qwen to DeepSeek V4.1 Flash.

We now serve DeepSeek V4.1 Flash on our own four NVIDIA RTX PRO 6000 Blackwell Server Edition GPUs in Hyderabad, India.

Endpoint: https://api.goedelmachines.com/v1
Model: deepseek/deepseek-v4.1-flash
Hub model: deepseek-ai/DeepSeek-V4.1-Flash
International USD list per million tokens: $0.30 input / $0.006 cached input / $1.20 output, including reasoning. Our India-only INR discount is separate.

The API supports OpenAI-compatible chat and streaming, usage reporting, reasoning, function calling, JSON object mode and JSON schema outputs. A bounded October 26 public-HTTPS smoke run passed 20/20 checks across our internal and dedicated partner keys. The dedicated partner keys report X-Data-Policy: none.

In three short rounds of 256 output tokens per request through the public endpoint from Colossus, median aggregate output divided by complete request wall time was 227 tokens/s at concurrency 1 and 458 tokens/s at concurrency 4. These are smoke measurements, not sustained-capacity or overseas-latency claims.

Our HF-style model list is https://api.goedelmachines.com/v1/models/hf. The cost endpoint is POST /v1/billing/costs; responses include Inference-Id. Nine of nine HF test requests were priced correctly, repeated lookups were identical, and another partner key could not access their costs. We have since deployed persisted per-request costs and verified that old request costs remain unchanged across the price change.

We have a dedicated HF integration key available privately. We would welcome the next onboarding step, including server-side registration, model mapping and client-integration requirements, before activation.

Organization: https://huggingface.co/GoedelMachines · Contact: rk@goedelmachines.com

Hi Hugging Face team,

We're IronStratum : we run a per-token, OpenAI-compatible inference API at https://api.ironstratum.com/v1.

We have several models exclusive to our platform, on GPU capacity we operate ourselves, alongside a catalog covering the DeepSeek, Qwen, GLM, and MiniMax families, with fresh builds listed fast (we listed
DeepSeek V4.1 Flash 1017 two days after its release) and specialty lanes for OCR, classification, voice, robotics and image generation. Several of these are Hub models with no inference provider today. They ship on dedicated action and vision routes with per-call and per-image metering.

We'd like to be listed as an Inference Provider. Following the registration guide, the pieces on our side are in place:

• Chat completions follow the OpenAI API shape
• /v1/models is anonymously readable and exposes pricing (USD per million input/output tokens) and context_length, in the format the guide expects.
• The billing endpoint is implemented to the documented batch protocol:
  POST with requestIds, responses in nano-USD integers, idempotent, inside the 30-minute window, authenticated by a dedicated key. The response header carrying our request id is x-request-id
• Tool calling and structured output (response_format with json_schema) are both live through our API and verified on our serving lanes.

Our organization on the Hub is iron-stratum (Team plan). We're ready to open the huggingface.js PR, register our initial mappings in staging, and follow with the huggingface_hub PR and a provider page for the docs.

What's the best way to start the huggingface.js PR first, or the server-side enablement with you first? Happy to share details on infrastructure, latency, or pricing.

Thanks

Hi Hugging Face team,

We are Apex Fusion (https://huggingface.co/apexfusion) and would like to be
registered as an Inference Provider. Conversational task only for now,
OpenAI-compatible API, served from our own GPU server in Helsinki, Finland.

• huggingface.js PR: https://github.com/huggingface/huggingface.js/pull/2686
  (check, lint and format pass; @Wauplin and @SBrandeis are tagged)
• Endpoint: https://hf-edge.vector.apexfusion.org (/v1/chat/completions,
  streaming, tool calling, structured output; /v1/models exposes pricing
  and context_length)
• Inference-Id response header and the billing endpoint are implemented
• First model: Qwen/Qwen3.8-27B

We will send the billing URL and the bearer token through a private channel
once you tell us where. Happy to adjust anything in the PR.

Thanks,
Apex Fusion