Spaces:

---

huggingface

/

HuggingDiscussions

like 94

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 !!

Hi HF team 👋 @julien-c

We'd like to register io.net (IO Intelligence) as an Inference Provider.

Provider ID: io-intelligence · Hub org: https://huggingface.co/ionet-official · API: https://api.intelligence.io.solutions/api/v1 (OpenAI-compatible)

io.net is an enterprise grade GPU network; IO Intelligence is our production inference API on top of it. We're already serving ~30 open models live — DeepSeek V4/R1, Llama 3.3/4, Qwen3.x, Kimi K2.x, GLM 4.x/5, gpt-oss, MiniMax — with streaming, tool calling, structured outputs, and reasoning.

Public GET /v1/models with per-model pricing, context window, precision/quantization, tool and reasoning support, and live latency/throughput stats — we publish the quantization each model is served at
Our model IDs already match HF Hub IDs (e.g. deepseek-ai/DeepSeek-R1-0528, meta-llama/Llama-3.3-70B-Instruct), so mappings will be 1:1

Hi @Wauplin @julien-c ! We would like to be listed as an Inference Provider on the Hub.

LLM Tech — an EU inference provider (registered in Poland, hardware in
Helsinki, Finland). We run open-weight models on dedicated single-tenant
GPUs and tune each deployment ourselves.

Current model: Qwen3.8-27B in NVFP4 on an RTX PRO 6000 Blackwell.
Measured on this deployment: 83 tok/s single-stream, 48 tok/s per request
at 32 concurrent, median TTFT 246 ms under sustained load, GPQA Diamond
78.3%. Endpoint: https://api.llmtech.eu/v1

What is already in place on our side:

• OpenAI-compatible chat completions with streaming, tool calling and
  structured outputs
• Inference-Id header on every response, including streaming
• Billing API returning per-request cost in nano-USD, idempotent
• /v1/models exposing pricing and context_length in the expected format
• Zero data retention, EU-only processing

Two questions before we go further:

1. How do payouts to providers work — schedule, method, any minimums?
2. Is there anything you would like to see beyond the integration steps in
   the docs before we open the huggingface.js and huggingface_hub PRs?

Artem Burei
LLM Tech — artem@llmtech.eu

Update, one day later: we are now live as a provider on NanoGPT.

LLM Tech is listed on their Qwen3.8-27B page and serving real user
traffic. Numbers from our own logs after the first day:

70 requests from 16 distinct client addresses
411k input tokens, 56k output tokens
0 failures on our side (one 400 from a malformed client request)
0.2s latency on their provider table — the fastest of 11 providers
measured TTFT 0.11s on our side, first token including reasoning

We are the only EU provider on that model, and the only one serving
NVFP4. Prefix caching is live and billed at a reduced cache-read rate.

Everything from my earlier message still stands and is now proven under
real traffic: OpenAI-compatible API, Inference-Id header on every
response, idempotent billing endpoint returning per-request cost in
nano-USD, /v1/models exposing pricing and context length.

Happy to open the huggingface.js and huggingface_hub PRs whenever you
give the word — the provider helper code is written and tested.

Artem Burei
LLM Tech — artem@llmtech.eu

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

We'd like to register JetInfer (https://huggingface.co/jetinfer) as an inference provider.

Who we are. JetInfer is an EU-based inference provider (legal entity: Quantive Bt., Hungary), founded by a low-level C++/systems engineer with ~21 years of experience. Our focus is serving open-weight models reliably at the lowest price.

What we serve today. Qwen/Qwen3.8-27B (conversational), OpenAI-compatible, streaming, tool calling and structured output verified live. $0.34 input / $2.55 output per 1M tokens, $0.034 cached input. 150k context. Zero data retention.

Integration status, per the register-as-a-provider guide:

• Task API: strictly OpenAI-compatible chat completions — live at https://api.jetinfer.com/v1 (catalogue: https://api.jetinfer.com/v1/models, health: https://api.jetinfer.com/health)
• Billing endpoint: implemented to spec (nano-USD, idempotent, batches up to 10k IDs) — ready to share the URL + auth privately
• Inference-Id header: on every response, including streaming
• /v1/models pricing + context_length: exposed in the documented format
• JS helper PR: https://github.com/huggingface/huggingface.js/pull/2402
• Python helper PR: https://github.com/huggingface/huggingface_hub/pull/4724
• Org is on the Team plan; ready for the model-mapping API once our account is enabled server-side

Could you enable our account server-side when you get a chance? Happy to provide anything else you need — test keys, SVG icon, docs page PR.

Thanks!
Richard Nemeth — JetInfer (https://jetinfer.com · https://x.com/jetinfer)

Hello Hugging Face Team,

We would like to add TextCLF as a Hugging Face inference provider.

Provider Details
Organization: TextCLF — https://huggingface.co/textclf-ai
Website: https://textclf.com

TextCLF provides low-cost LLM inference powered by our proprietary TQ 4-bit quantization and CUDA inference stack. A key differentiator of TQ is that it is completely calibration-data-free: it requires no calibration dataset, so the quantization is not optimized around or dependent on the distribution of a particular calibration set. This makes TQ particularly attractive for generalization across domains and workloads beyond a calibration dataset, while achieving fidelity competitive with leading quantization methods.

On Qwen 3.8 27B (WikiText), TQ 4-bit achieves 0.0282 mean KL divergence and 92.4% Top-1% agreement versus the full-precision model, with approximately 4.1× model compression.

Our currently available models and pricing per 1M tokens are:

Llama 3.1 8B Instruct: $0.018 input / $0.038 output
Llama 3.3 70B Instruct: $0.08 input / $0.30 output
Qwen 3.8 27B: $0.30 input / $2.50 output ($0.03 cached input)

Coming soon: DeepSeek V4 Flash 0731 and XiaomiMiMo MiMo V2.5.

We have submitted PRs to both Hugging Face client repositories to add TextCLF provider support:

Python: https://github.com/huggingface/huggingface_hub/pull/3895
JavaScript: https://github.com/huggingface/huggingface.js/pull/2022

Could you please enable the Model Mapping API for our organization and provide guidance on registering TextCLF as an inference provider?

Thanks,
TextCLF Team
Email: contact@textclf.com

Hi! Following up on our provider application for LLM Tech:

• huggingface.js PR: huggingface/huggingface.js#2403
• huggingface_hub PR: huggingface/huggingface_hub#4727 (full test file passing)

Our endpoint has been serving live production traffic since Aug 22 (currently via NanoGPT): 100% uptime, zero failed server requests, and the best measured TTFT among the 12 providers listed for this model there. Live status with measured numbers, refreshed every 5 minutes: llmtech.eu/status

Specs: Qwen3.8-27B (NVFP4 on Blackwell), 262,144-token context, streaming, prompt caching billed at $0.04/M. Zero data retention, EU jurisdiction.

The billing endpoint per the provider spec is implemented and tested. The org is set up (huggingface.co/llmtech); we'll add the Team subscription and register the model mapping as soon as you're ready on your side.

Dear Hugging Face team,

we would like to register Gridmetry as an Inference Provider.

We operate production inference infrastructure in a Tier III+ datacenter in Munich, Germany and expose an OpenAI-compatible API.

Our initial model portfolio includes high-demand open-weight models such as:

• Kimi K3
• DeepSeek V4 Pro 0813
• MiniMax M2.7
• gpt-oss-120B
• Gemma-family models

We are specifically focused on providing high-throughput, cost-efficient dedicated inference capacity for production workloads in Europe.
We would like to add an initial capacity of ~20.18 billion tokens/day

Current benchmark examples include:

gpt-oss-120B: ~713 output tok/s
MiniMax M2.7: ~428 output tok/s / ~690 ms TTFT

Infrastructure is hosted under German/EU jurisdiction with ISO 27001:2022, zero-data-retention support and dedicated production capacity.

Hugging Face organization: Gridmetry GmbH

We would be happy to implement the required billing endpoint, Inference-Id handling, /v1/models pricing metadata and Hugging Face client integrations.

Could you please enable us for provider onboarding / model mappings and let us know the next steps?

Daniel Schade
COO, Gridmetry GmbH
dsc@gridmetry.com