Hugging Face: the open-model hub
Hugging Face: Open Models, Inference Providers and Endpoint Costs
How the Hugging Face Hub, Inference Providers and dedicated Inference Endpoints fit together, what popular open models cost per million tokens when hosted by third parties, and when to run them yourself.
Prices and versions: figures are published list prices in US dollars per 1 million tokens, read from each provider’s own pages on 18 November 2026. Models, prices and promotions change often, so confirm on the provider’s page before you budget.
What Hugging Face gives you
- The Hub. The main public home for open-weight models, datasets and demos. The models page we read shows over three million models.
- Inference Providers. One OpenAI-compatible API that routes to partner hosts such as DeepInfra, Together, Novita, Nscale and others, billed through your Hugging Face account.
- Inference Endpoints. Dedicated, autoscaling deployments on GPUs you rent by the hour, for steady traffic or models that serverless hosts do not offer.
- Spaces and Jobs. Hosted demos and batch compute, which can draw on the same credits.
Open models and what hosts charge
List prices per million tokens for popular chat models on the Inference Providers page. Each model is offered by several hosts at different prices, speeds and context limits, and we show the lowest priced host we could read. Prices change frequently.
| Model | Cheapest listed host | Input | Output | Context |
|---|---|---|---|---|
| gpt-oss-20b | deepinfra | $0.03 | $0.14 | 131K |
| gpt-oss-120b | deepinfra | $0.04 | $0.17 | 131K |
| Llama 3.1 8B Instruct | novita / deepinfra | $0.02 | $0.05 | 16K to 131K |
| Qwen3.5-9B | deepinfra | $0.10 | $0.15 | 262K |
| Gemma 4 31B it | novita | $0.14 | $0.40 | 262K |
| GLM-5.3-Flash | several hosts | $0.15 | $0.50 | 1.05M |
| MiniMax-M3.1-Flash | novita | $0.15 | $0.60 | 1M |
| DeepSeek-V4.1-Flash-1017 | deepinfra | $0.20 | $0.60 | 1.05M |
| Qwen3.8-27B | nscale / deepinfra | $0.20 | $2.50 | 262K to 1.05M |
| GLM-5.3 | deepinfra | $0.90 | $4.00 | 1.05M |
| Kimi-K3.1-Pro | together | $2.00 | $10.00 | 1.05M |
| Kimi-K3 | deepinfra | $2.85 | $14.25 | 1.05M |
Hugging Face says it charges the same rate as the provider with no additional fees. The trending list on the Hub at the time we read it was led by Qwen3.8-27B, DeepSeek-V4.1-Flash-1017 and Kimi-K3.1-Pro, plus many quantised and fine-tuned derivatives.
A monthly bill example
For 100 million input and 20 million output tokens a month, gpt-oss-120b at the listed rate costs about $7.40, GLM-5.3-Flash about $25, DeepSeek-V4.1-Flash-1017 about $32 and GLM-5.3 about $170. That is one to two orders of magnitude below frontier closed models, though quality, safety behaviour and tool use differ, so evaluate before switching. Compare with the OpenAI and Anthropic examples.
Plans and credits
Inference Providers is a paid service with small monthly credits for experiments.
| Account | Price | Monthly inference credits |
|---|---|---|
| Free | $0 | None; pay as you go after buying credits |
| PRO | $9 / month | $2 |
| Team | $20 / user / month | $2 per seat, shared |
| Enterprise | From $50 / user / month | $2 per seat, shared |
Team and Enterprise usage is billed to the organisation only when the request names it, and organisations can cap spend and disable providers.
Dedicated endpoints and GPUs
When traffic is steady, privacy matters or the model is not offered by a host, a dedicated endpoint can make sense. Hourly prices for selected AWS GPU instances:
| Dedicated GPU (AWS) | Memory | Per hour | Per month |
|---|---|---|---|
| NVIDIA T4 | 14GB | $0.50 | $360 |
| NVIDIA L4 | 24GB | $0.80 | $576 |
| NVIDIA L40S | 48GB | $1.80 | $1,296 |
| NVIDIA A100 | 80GB | $2.50 | $1,800 |
| NVIDIA H200 | 141GB | $5.00 | $3,600 |
| NVIDIA B200 | 179GB | $9.25 | $6,660 |
CPU instances start at about $0.03 an hour. Dedicated capacity costs the same whether or not it is busy, so it pays off at high, steady utilisation. Past a certain size, renting neocloud capacity or running your own GPUs becomes cheaper again.
Hosted, dedicated or self-run
- Serverless providers. Lowest effort and lowest cost for spiky or experimental use. You share capacity and accept each host’s logging and availability.
- Dedicated endpoints. Predictable latency and isolation, billed hourly.
- Self-hosted. Full control for sensitive data and very high volume. Needs GPUs, a serving stack and operations, which is where we help.
- Licence check. Open weights come with different licences. Review commercial-use, attribution and acceptable-use terms before you ship.
Where this meets infrastructure
We help teams pick open models, size the hardware behind them and stand up serving: open-source AI models, hosted AI models and the application layer. Gateways such as OpenRouter can sit alongside Hugging Face endpoints, and the agent platform and security controls wrap the whole.
Where this fits in the stack
Common questions
Is Hugging Face free?
Browsing and downloading from the Hub is free. Inference Providers is a paid pay-as-you-go service with small monthly credits (none on the free plan, $2 on PRO and $2 per seat on Team and Enterprise).
Does Hugging Face add a markup to inference prices?
It says it charges the same rates as the provider with no additional fees, so the price depends on the host you are routed to.
When should I use an Inference Endpoint instead of serverless?
When traffic is steady, latency or isolation matters, or the model is not offered by a serverless host. For spiky or experimental use, serverless providers are usually cheaper.
Can I use these models commercially?
It depends on each model’s licence. Check commercial-use and acceptable-use terms on the model card before you ship.
Sources and dating
Source: Hugging Face Inference Providers, pricing and Hub pages, read 18 November 2026. Part of the Inference Providers list was summarised from a truncated page, so treat per-model rows as indicative and check the live page. Some figures are vendor claims or preliminary and some differ between sources, so we confirm against the provider pages before we size or budget a build. Hugging Face and the model names are trademarks of their owners. Embedica is independent and is not affiliated with or endorsed by any of the companies named.
Next steps
Related: LLM models and token costs · OpenAI models · Anthropic Claude models · Neoclouds and GPU cloud
Deciding between hosted, dedicated and self-run open models?
Open model assessment
