---
title: "Quickstart: web"
sidebar_label: Web
description: "From a fresh account to a benchmark result and a running endpoint in the browser, with the minutes each step takes."
---

Time to first result on a two-core Linux laptop: about 25 minutes, 15 of them waiting for the run. The [tutorial](../first-benchmark/01-overview.md) explains each step; this page is the short form.

1. **Sign in.** From a signup code or an invitation ([Sign up with a signup code](../sign-up-with-a-signup-code.md)). The verification link lands you on the **Default Project** dashboard.
2. **Register a device** (2 minutes). **Devices**, **Register Device**, the tab for the machine's operating system, copy the command, run it on the machine. Leave the **Reusable, many devices** switch alone for now: turning it on asks whether to keep the command you copied or make a new one, and a new one revokes the one you copied. The dialog turns into the device list when the device arrives, within a minute. Give the device one more minute before step 4.
3. **Add a model** (1 minute). **Models**, **Add Model**, paste a Hugging Face URL or `owner/repo`; `HuggingFaceTB/SmolLM2-135M-Instruct` is a good first model. A model somebody registered before you comes back as a shared record under your models; your results on it are yours.
4. **Run a benchmark** (15 minutes). **Benchmarks**, **New Benchmark**, pick the device and the model, **Continue**, then **LLM Performance** with **Quick** on, **Run Benchmarks**. The log shows the engine push for the first five minutes and then one line per measured cell; the progress bar stays at 0 of 1 until the leg completes.
5. **Read the result** (1 minute). The run's page: TTFT, Tokens/sec, Peak decode, TPOT and Peak Memory, each with median, minimum and maximum, and a table of cells (`128:128` is prompt tokens : output tokens, `c8` is eight parallel streams). **Share Results** makes a link.
6. **Deploy it** (2 to 3 minutes after a benchmark on the same device; 5 to 9 on a device that has never run the engine). **Deploy** on the result, keep the name, **Deploy**. The page reads "Starting" and "Waiting for logs" while the engine starts on the device, then **Running** with the platform address, a hidden key (**Reveal key**) and the device's own address. Send one request:

```bash
curl -s -X POST "https://<platform>/api/v1/servings/<serving>/devices/<device>/proxy/v1/chat/completions" -H "Authorization: Bearer <deployment key>" -H "Content-Type: application/json" -d '{"model":"<id from /v1/models>","messages":[{"role":"user","content":"Hello"}]}'
```

7. **Invite your team.** **Settings**, then **Organization**, **Add member**: the organization role is Admin or Member; give project roles (Developer, Viewer) on the project's **Members** tab. Only an owner or admin can open a device's terminal. [Roles](../../concepts/roles.md) has the table.

What spends credits, where your plan meters them: every benchmark leg and every deployment start sends the engine to the device, and the **Subscription** card under Settings shows the balance.