Spaces:

---

huggingface

/

HuggingDiscussions

Like 97

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 !!

Hello Hugging Face team,
We’d like to add Telnyx as a Hugging Face Inference Provider.

Provider details:

1. Organization: Telnyx
2. Website: https://telnyx.com
3. Inference docs: https://developers.telnyx.com/docs/inference/getting-started/index
4. Models: https://developers.telnyx.com/docs/inference/models
5. Pricing: https://telnyx.com/pricing/inference-api

Telnyx provides OpenAI-compatible inference on Telnyx-operated GPU infrastructure, with support for chat completions, streaming, function calling, structured outputs/JSON mode, and embeddings.

We’d like to start with conversational model mappings for Telnyx-hosted open-weight LLMs, then add additional supported tasks as appropriate.
Could you advise on the preferred next step for adding Telnyx as a provider and enabling the Model Mapping API for the Telnyx Hugging Face organization?

Thanks,
Telnyx Team

Hi HF team (cc @Wauplin @SBrandeis @julien-c @hanouticelina),

I would like to register BroadNet as an inference provider.

We're a small shop: Dubai HQ, our own GPUs in USA, Jordan, UAE, KSA, and we've been running Arabic-heavy
production for our own customers for a while. Things we bring that are thin on the Hub
right now:

• Arabic speech-to-text that's actually serveable. CohereLabs/cohere-transcribe-arabic-07-2026
  has no provider on the Hub today. It's Apache-2.0 and sits at #2 of 37 on the Open Universal
  Arabic ASR Leaderboard (#1 is under a licence that doesn't allow serving it as an API). We also
  run the multilingual CohereLabs/cohere-transcribe-03-2026. Both measured at ~45x realtime
  single-stream and ~190x batched on Arabic audio, on our hardware.
• Qwen3-VL instruct models with vision: Qwen/Qwen3-VL-30B-A3B-Instruct (int4, pooled
  across three boxes) and Qwen/Qwen3-VL-8B-Instruct (int8, 78 tok/s single-stream).
• Custom Arabic ASR and Vision models: our internal battle tested fie tuned OSS models and LoRAs.

What's ready:

• Provider id broadnet, Hub org https://huggingface.co/Broadnet
• OpenAI-compatible base URL for HF traffic: https://inference.broadnet.ai/hf — chat with
  streaming, tool calling and JSON-schema structured output, and /v1/audio/transcriptions for ASR
• GET /v1/models on that base with pricing (USD per M) and context_length
• Inference-Id header on every response, streams included
• Billing endpoint (nano-USD per request id) at POST /hf/billing/costs, same bearer as inference —
  happy to share a test key privately
• huggingface.js PR: https://github.com/huggingface/huggingface.js/pull/2445
• Zero data retention, written up at https://broadnet.ai/inference-privacy

Prices: 30B-A3B $0.13 / $0.52 per M, 8B $0.117 / $0.455 per M, Arabic STT $0.0042/min,
multilingual STT $0.002/min.

Could you enable the org server-side so we can register the mappings? We'll start everything in
staging and flip to live once your monitor is happy. Logos are ready whenever you want them.

Thanks!
MJ

Hi Hugging Face team,

We’d like to register Bridge as an Inference Provider on the Hub.

Bridge operates cost-efficient OpenAI-compatible LLM inference infrastructure in Asia. Our initial integration would focus on Qwen3.8-27B.

Current production capability:

• OpenAI-compatible /v1/chat/completions
• Streaming SSE
• Streaming and non-streaming usage reporting
• Tool/function calling and automatic tool selection
• Structured JSON outputs
• Authenticated HTTPS API
• Text-only inference
• 70K context / 8K max output
• $0.15/M input / $1.98/M output
• ~651 aggregate output tokens/sec verified at 16 concurrent requests

Website:
https://chyyx.github.io/Bridge-ai/

Privacy Policy:
https://chyyx.github.io/Bridge-ai/privacy.html

We have reviewed the Inference Provider registration guide and are ready to proceed with the huggingface.js integration and the Hugging Face-specific model mapping and billing requirements.

Proposed provider ID: bridge

Could you please confirm whether we should proceed with the huggingface.js PR now, and when you would like us to enable the required Hub organization plan for Model Mapping API access?

Thanks,
JC
Bridge

Hi Hugging Face team!

We’d like to start the process of registering Netra Runtime as a Hugging Face Inference Provider.

• Name: Netra Runtime
• Proposed provider identifier: netra
• Website: https://netraruntime.com
• API base URL: https://api.netraruntime.com/v1
• API documentation: https://app.netraruntime.com/docs
• API reference: https://api.netraruntime.com/docs
• OpenAPI schema: https://api.netraruntime.com/openapi.yaml
• API compatibility: OpenAI-compatible /v1/chat/completions and /v1/models
• Authentication: Bearer API keys
• Streaming: SSE streaming with [DONE]
• Request identification: Unique X-Request-Id response header
• Usage accounting: Input, cached-input, reasoning, and output-token usage

Netra Runtime provides optimized inference for open-weight models across AMD and NVIDIA GPUs. Our focus is high sustained throughput, competitive token pricing, and reliable OpenAI-compatible inference.

Our currently published model is:

• deepseek/deepseek-v4-flash-0731
• Uncached input: $0.20 per million tokens
• Cached input: $0.05 per million tokens
• Output: $0.50 per million tokens
• Supports streaming and configurable reasoning levels

Could you please confirm:

1. Whether netra is an acceptable provider identifier.
2. Whether onboarding can begin with chat completions, streaming, and reasoning while tool calling remains unsupported for the initial model.
3. Which Hugging Face model mapping you would prefer for our initial DeepSeek route.
4. Whether you would like us to open the huggingface.js provider PR before server-side enablement.
5. The preferred private channel for reviewer credentials and future billing-endpoint credentials.

Once the expected capability scope and initial mapping are confirmed, we can prepare the client integration and remaining provider-specific billing work according to the registration guide.

Thank you!

Farrel
https://netraruntime.com

Hi Hugging Face team,

Following up on our Bridge provider introduction above. Could you advise whether we should proceed directly with a huggingface.js integration PR for the proposed provider ID bridge, or whether your team needs to review our provider eligibility first?

We would also appreciate guidance on the preferred private channel for sharing a test API key and any additional information needed for evaluation.

Our initial integration would focus on Qwen3.8-27B INT4, text-only. We’re happy to follow your preferred onboarding sequence.

Thanks,
JC — Bridge

Hi — we'd like to be listed as an Inference Provider on the Hub.

Heabsy (FEYA s.r.o., Slovakia) serves open models on dedicated GPUs behind an OpenAI-compatible API. Three models today, all NVFP4 with the full 262k context:

Inference is zero-retention: prompts and completions are never written to storage — the response cache is switched off in our gateway precisely so that claim is literally true, and we can demonstrate it on our production database rather than just assert it in a policy.

On the prerequisites in the registration guide: our API is strictly OpenAI-compatible, SSE streaming carries usage with the cached-token split, tool calling and JSON-schema structured output both work, reasoning is returned in a separate field, and every response — streaming included — already carries a unique per-request id header, so billing reconciliation has something to key on. Public machine-readable catalogue: https://api.heabsy.com/openrouter/v1/models

We're ready to open the huggingface.js PR and stand up the billing endpoint described in the guide. Two questions before we start: who should we coordinate with, and is there anything you'd want to see first? Happy to hand over a test key.

Konstantin Fomichev — Heabsy / FEYA s.r.o.

Hi Hugging Face team (cc @julien-c @Wauplin @SBrandeis @hanouticelina)

We'd like to register SolarGPU as an Inference Provider on the Hub.

• Name: SolarGPU, proposed provider id solargpu
• Hub account: https://huggingface.co/solargpu (organization with Team plan to follow at step 3)
• Website: https://solargpu.es · privacy & retention: https://solargpu.es/privacidad · terms: https://solargpu.es/terminos-api
• Who we are: a Spanish company (Acceleration Robotics S.L.U.) running solar-powered GPU nodes in Spain. We serve open-weight models from our own NVIDIA RTX PRO 6000 Blackwell (96 GB) cards — one dedicated today, expandable to five across two nodes (480 GB VRAM) — with EU data residency and zero data retention (prompts and completions are not stored; billing metadata only; no training on customer data).
• API: OpenAI-compatible /v1/chat/completions with streaming, usage on every response, tool calling and structured outputs; served with vLLM 0.29 behind a LiteLLM gateway. Hub traffic would enter at https://api.solargpu.es/v1 (live; https://api.solargpu.es/models publishes the catalogue with pricing and context length).
• Initial mapping, conversational: openai/gpt-oss-120b → gpt-oss-120b, Qwen/Qwen3.6-35B-A3B → qwen3.6-35b-a3b, nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-BF16 → nemotron-3-super-120b (nvfp4), NousResearch/Hermes-4-70B → hermes-4-70b (fp8); proposed in staging. On the full node we can add deepseek-ai/DeepSeek-V4-Flash-0731 and zai-org/GLM-5.3-Flash.
• Pricing: we keep our list prices at or below the current floor for these models (USD per M tokens, input / output): gpt-oss-120b 0.03 / 0.17, Qwen3.6-35B-A3B 0.05 / 0.70, Nemotron 3 Super 0.08 / 0.38, Hermes 4 70B 0.12 / 0.38. /v1/models will expose pricing and context_length as required.
• Ready on our side, per the register-as-a-provider guide:
  • Task API OpenAI-compatible; TTFT well under 5 s at our concurrency; tool calling and structured output supported (vLLM).
  • Billing endpoint to your spec (cost in nano-USD, polled per minute) — implemented with the test endpoint; URL and auth to follow privately.
  • huggingface.js provider helper — pull request to follow once you confirm the id.
• Timeline: the endpoint is live now; a test key for your validation suite on request; 1 GPU dedicated at launch, up to 5 on demand.

Why us: most floor-price endpoints for open-weight models are outside the EU. We offer EU residency with zero retention at floor prices, from solar-powered hardware in Spain.

Happy to share a test key so your team can run the validation suite against the endpoint, and to discuss the Team plan for the org as part of step 3.

Thanks!
SolarGPU team