LLM Launcher — Catalog

Importable launch recipes for llml. Stop reconstructing llama-server flags from shell history — grab a profile someone already tuned for your model and hardware, import it, and run.

Browse the catalog →

Note

The catalog is growing. Alongside community PRs, the maintainers regularly find and add new profiles. Watch this repo (Watch → Custom → Releases) to get notified when new profiles land — or check back any time.

A profile is a TOML file that captures the exact args and env vars someone converged on for a specific model, backend, and hardware configuration. Export yours from llml, share it here, and others with the same setup can import it instead of starting from scratch.

Profile format

Profiles use schema_version = 3. A minimal profile looks like:

schema_version = 3

[[profiles]]
name = "llama-3-8b-q4-rtx4090"
backend = "llama"
model_hint = "Meta-Llama-3-8B-Instruct.Q4_K_M.gguf"
args = [ "--n-gpu-layers 99", "--ctx-size 8192" ]

[profiles.use_case]
primary = [ "general" ]
tags = [ "interactive" ]

[profiles.hardware]
class = "gpu"
min_vram_gb = 24
notes = "Tested on RTX 4090, CUDA 12.4, Ubuntu 24.04"

The full schema is defined in llml/docs/profile-format.md.

Adding a profile

1. Write a .toml file following the format above. Use an existing profile as a reference.
2. Place it in profiles/. Name it after the model and quant (e.g. llama-3-8b-q4-rtx4090.toml).
3. Open a PR. CI validates the schema before merge.

That's it. No accounts, no signup — just a TOML file and a pull request.

Before opening a PR, check that:

• backend, model_hint, and hardware.class are set
• use_case.primary, when set, uses general or eval
• hardware.min_vram_gb matches what your config actually uses
• hardware.notes names the real machine you tested on
• The filename is descriptive

Importing a profile

# Import from the catalog URL
llml import https://llml.dev/profiles/Qwen3.6-enable-thinking.toml --activate

The profile appears under p in the llml TUI on next launch.

Repository layout

profiles/           one .toml file per profile — PRs add new profiles here
site/               static catalog site (Astro), deployed to GitHub Pages
design/             design system: colors, type, components
scripts/            scraping pipeline and dev-docs utilities
.github/workflows/  CI: validate incoming profiles, build + deploy site

Scraping pipeline

An automated pipeline extracts profiles from model card sites and writes portable TOML files. See scripts/README.md for setup, CLI flags, and provider configuration.

License

Profiles are released under CC0 1.0 — public domain, no attribution required. Site source and design assets are MIT.