Run an Open Model on a Rented GPU for a Day: An Endpoint Your Tools Can Call, GPU Hours by the Minute, and the Break-Even Against the API

A small team wants an open model of its own for a day: a private endpoint that its scripts can call the way they call an API, running a model whose weights it chose, on a GPU it rents by the minute and switches off at six. This page builds that endpoint two ways, in a browser on Hugging Face Inference Endpoints and with one command on Together, then does the arithmetic that decides whether to keep it running: the GPU by the hour times 730, against the API by the token. The steps come from the hosts' own documentation, including one place where two pages of the same documentation disagree.

What you hand over

An HTTPS endpoint, private, behind a token, serving a named open-weight model, that answers the same chat request your code already sends to a hosted API. A note of what it cost for the day, to the minute. And a stopped deployment at the end of it, so tomorrow's bill is zero until you start it again.

Step one: the licence on the weights

The GPU costs the same whatever runs on it, so the first read is the model's licence, not the price list. Mistral's pricing page says its open-weight models are Apache 2.0 licensed for research and individual use, while commercial deployments require a Mistral license with separate terms for derivatives and production use, and that you can self-host them anywhere. Every model card on Hugging Face carries its own licence line, and the catalog described below does not read it for you. If the output will be sold, that line decides before the first minute is billed.

Path A: in the browser, from the Catalog

Hugging Face's quick start is a browser walk. Inference Endpoints is available to accounts with a card on file, so add one in billing settings first; the guide recommends automatic recharge. Then open the Inference Endpoints UI and click Catalog. The catalog holds popular models with tuned configurations for one-click deploys, filterable by name, task and hardware price. The guide searches for Llama 3.2 3B Instruct, opens the card, and keeps every pre-selected option, noting that the Nvidia L4 is the recommended hardware for that model. Click Create Endpoint.

Initialising takes about 3 to 5 minutes. When the endpoint reaches running, the Overview tab shows the Endpoint URL, which is what your code calls. The Playground tab is a quick visual test, and its API section gives a code snippet you paste an access token into, because all endpoints are created private and every request is authenticated and encrypted in transit. The guide ends with the three things you can do afterwards: pause the endpoint, delete it, or leave it, in which case it scales to zero on its own.

Path B: one command on Together

Together AI's dedicated model inference is the same result from a terminal. With the Together CLI installed and your API key exported, one command deploys a supported model:

tg beta endpoints deploy google/gemma-4-E4B-it --endpoint quickstart-endpoint

It creates an endpoint, attaches a deployment on the model's default hardware, routes all traffic to it, and prints the endpoint string in the form project-slug/quickstart-endpoint. First-time provisioning takes about 5 to 10 minutes for a model that size while the weights download and hardware is allocated; tg beta endpoints get with the deployment id reports when it reaches ready. The console does the same from Endpoints, New endpoint: an endpoint name, a deployment name, the model, quantization, hardware, region, and min and max replicas, both defaulting to 1.

Pointing your tools at it

Together's design point is that the request does not change. Set the base URL to api-inference.together.ai/v1 and pass the endpoint string as the model parameter of the ordinary chat completions call, with the same messages and max_tokens you send to a serverless model; the docs say you can prototype on serverless and move to dedicated without changing application code. On Hugging Face, the Playground's snippet does the same for the Endpoint URL with your token. Run the day's job: the batch of documents, the evaluation set, the internal tool.

The day by the minute

Inference Endpoints prints hourly rates and bills by the minute, charging while an endpoint is initialising and while it is running: Nvidia T4 $0.5 an hour, L4 $0.8, A10G $1, L40S $1.8, A100 80 GB $2.5, H200 $5, with CPU instances from $0.033. Its formula is the hourly rate times hours times minimum replicas. An eight-hour working day on the guide's L4 is $6.40; the same day on an A100 is $20. Together bills per minute per running replica at its hourly rates: H100 $3.99 on demand or $1.99 preemptible, H200 $5.99, B200 $8.99. fal rents an H100 for your own deployments from $1.89 an hour. The AI job cost calculator multiplies any of these by the hours you type.

The month in hours

Now the number that decides whether the endpoint stays up. The chart above multiplies each hourly rate by 730, a month of hours. The small T4 on Hugging Face's hardware price list, $0.40 an hour, is $292; an Inference Endpoints L4 is $584. fal's H100 is $1,380, Together's preemptible H100 $1,453, its on-demand H100 $2,913, its B200 $6,563, and eight A100s on Hugging Face $14,600. Those are the prices of a GPU that never sleeps; an idle GPU costs the same per minute as a busy one until it is scaled to zero, which is the next section.

The same month, in tokens

Put the API next to it. Mistral Large is $0.5 per million tokens in and $1.5 out, with batch processing at half price. $2,913, the on-demand H100 month, buys about 1.9 billion output tokens at that rate. Together's serverless prices are lower still: DeepSeek V4 Flash at $0.14 in and $0.28 out, MiniMax M3 at $0.30 and $1.20, Kimi K3 at $3.00 and $15.00. A team that will not push a billion tokens a month, and does not need the model, the data or the latency to sit on one machine, is paying for an empty GPU. A team that does, or that runs a fine-tune no API serves, has its answer in the other column. The drafting page has the hosted API price lists in full.

Stopping it, and the two idle clocks

On Together, a deployment runs until you stop it. Stop, in the console or by setting both min and max replicas to 0, scales it to zero replicas and releases the hardware; once it reaches the STOPPED state it is neither billing nor serving, and Start brings it back with the replica counts you confirm. Deletion is permanent and ordered: stop the deployment, delete it, then delete the endpoint once it has no deployments. The quick start's own last step is tg beta endpoints rm with the endpoint id and --force.

On Hugging Face, the quick start says the endpoint scales down to zero by default, becoming idle after 1 hour of inactivity. The autoscaling guide, describing the Automatic Scale-to-Zero setting, says an endpoint with no requests for over 15 minutes is scaled to 0 replicas. Two pages, two idle periods, for a GPU billed by the minute: the 45 minutes between them is 60 cents on an L4 and $3.75 on an H200, per idle stretch. The autoscaling guide also spells out the price of zero: a new request triggers a cold start whose length depends on the model's size, the server answers 502 Bad Gateway while the replica initialises, and there is no request queue, so your client must retry. Check the Settings page of your own endpoint for the period actually applied, then pause it by hand when the day is done rather than trusting either page.

The clauses that bite a small team

Together is pay as you go with non-refundable fees; credits funded by ACH or wire are not usable until they clear, access can be suspended on notice if any amount is 5 days past due, and you may not terminate the agreement without Together's written consent. Hetzner, if the team prefers a dedicated box without a GPU, prices the AX42-1 at EUR 97.30 a month plus a one-off EUR 49.00 setup fee, net of 19% VAT for German customers; backups are the customer's duty and there is no claim to a particular IP address.

What leaves with you

Mistral's open weights can be self-hosted anywhere, so a workflow built on them does not depend on any one host staying at any one price. Hugging Face says you own the Content you create and will not sell it. DigitalOcean's docs say snapshots and backups cannot be downloaded off the platform; getting data out means copying it yourself with rsync or SFTP. Keep the weights, the system prompts and the data on a disk you control.

What ends it

Hugging Face reserves the right to suspend or terminate access at any time, with or without cause, with or without notice, and may remove your Content at its sole discretion if it has a concern about it. Together's agreement cannot be ended by you without its written consent. DigitalOcean can terminate access at any time, effective immediately, and deletes Services Content on termination under its standard closure practice. For a team whose tool depends on one endpoint, the answer is the same as for the support bot: keep the pieces in your own folder and know the second host's deploy command.

What to do today

Put a card on file, deploy the catalog's Llama 3.2 3B on an L4 with the defaults, and run one real job through the Endpoint URL from the Playground snippet. Note the minutes on the billing page, then pause it. The day will have cost a few dollars; the number to write down is your team's real token volume for the month, because that, against $584 to $2,913, is the only figure on this page that decides anything.