The L80 Console, Screen by Screen

If you run AI models, you want to know how fast your server is, how good its answers are, and whether the weights you downloaded are what you think they are. Then you want to share that with other people in a form they can check. L80 is the launch80 toolbox for that, and L80 tui puts all of it in one terminal console.

Everything in the video is a real recording on a real machine, with no mockups. Long runs are fast-forwarded, and the video says so when they are. Nothing is published or rented in it: every publish and rent dialog is cancelled.

Install and start

curl -fsSL https://www.launch80.com/install.sh | sh
L80 tui

Every tool is free and runs on your machine. Only publishing a page needs a launch80 token, which you get as a Supporter. The first time you open the console, the onboarding tour checks your machine, sets up publishing and asks what brought you here, then drops you on the screen that does it.

Every screen, with a link to it

Getting around ▶ 0:33

The sidebar lists every screen: Home first, the rest alphabetical. The arrow keys (or j and k) move through it and preview each screen on the right. Enter opens a screen, escape goes back, tab moves to the next field, slash filters a list, and ? opens help from anywhere. q on the sidebar, or control-C, quits. Everything in the console is also a plain command, so you can script it or hand it to an agent.

Bench Speed: how fast is my server? ▶ 5:15

Give it the endpoint's base URL and the model name, and BetterBench measures it. In the video that's Qwen3 0.6B on llama-swap. Quick run does five passes per category instead of twenty, so you get a result in minutes. Phases picks decode, prefill and concurrency; All gives the fullest launch80 page. Notes are key=value pairs, like the engine and the hardware, shown on the published page. An endpoint API key is passed to BetterBench and never recorded. The endpoint address never leaves your machine.

The result is saved as JSON and an HTML report in the current directory, and p publishes it. L80 asks before anything goes out, and cancelling loses nothing: press p again whenever you're ready.

Bench Quality: is the answer still good? ▶ 7:28

Bench Quality runs QualityBench: math, instruction following, knowledge, code, tool calling, long context and more. Run measures one model. Compare takes two finished runs, such as full precision and a quantized copy, and reports the quality lost.

• Scope: Full runs every task (a few hours on a big model), Quick runs a fixed tenth of each, and Custom lets you tick tasks, each with its size and a time estimate. The video keeps only IFEval for a short run.
• Target: your own GPUs, or a rented Vast.ai box. L80 finds servers on this machine's usual ports by itself; in the video it found llama-swap serving twenty models.
• Compare to is optional: pick an earlier run of the same model and this run is also scored against it.

Before starting, L80 shows how long it expects the run to take and asks whether to go ahead. Results land in one QualityBench folder, so unpublished runs show up in Bench History from any directory.

Bench Risk: are these the weights I think they are? ▶ 9:46

Bench Risk checks a local directory, a Hugging Face repo or a served model for tampering, with the Poisoned Weights Detector. Pick a level:

Base and parent models are read from the model card if you leave them empty. The verdict comes first, then coverage and every check with what it observed.

History and Verify: results other people can trust ▶ 11:32

Bench History lists everything your key has published, newest first, with each page's trust tier. Unverified means self-reported. Verified means a launch80 verifier reproduced or reviewed it. Select a page and press v to ask for verification. Bench Verify is the other side: the queue of runs waiting for a check, and the status of your own requests. Scans and quick runs need a reproduction, rerun on the verifier's own hardware and compared.

Rent GPUs: no hardware needed ▶ 13:14

Rent GPUs rents a Vast.ai GPU, serves a model on it, runs your benchmarks in order, publishes the results and tears the box down. The four-page wizard covers the model and engine (vLLM or llama.cpp), the hardware and a price ceiling, the operations (BetterBench, then QualityBench), and a final review with a hard cap in hours and a budget.

• Searching offers is free and rents nothing.
• Dry run is on by default: it shows the plan without renting.
• When the run ends or hits the cap, L80 destroys the box. The Runs tab shows every rental and its page.

And the rest ▶ 15:21

Dashboard is live telemetry for your GPUs and inference servers, in the terminal or as a web page. Tools shows the pinned version of every tool and how it resolves, and an environment variable like L80_BETTERBENCH points L80 at your own copy. Templates lists the page types launch80 accepts, and --dry-run checks a payload you wrote yourself without publishing it. Skills installs L80's skills for your coding agent, so it can benchmark, rent, scan and publish for you; for agents without skills, paste L80 skills agents-md into AGENTS.md or CLAUDE.md.

Go further

• The console docs and the full command list
• github.com/launch80/L80-CLI
• The journey of a token: what the numbers Bench Speed reports actually measure

Recorded with L80 1.3.3. Narration: Kokoro-82M (voice af_heart).