New profiles added regularly — watch the llml repo ↗ to get notified.
llama-server flags from shell history.
llml finds your local models, detects your runtimes, and launches them with a saved profile.
Pick a model, pick a profile, press R.
Works alongside llama.cpp, Ollama, vLLM, and KoboldCpp. Your backend stays exactly as it is.
llml finds every local model — GGUF files, safetensors, Hugging Face cache — and lists the runtimes on your machine.
Choose a model, a runtime, and a saved profile. The generated launch command is shown before execution — no surprises.
Press R. One keypress, and the right command runs against the right model on the right backend.
Import the exact config someone already tuned for your model and hardware — one command, it runs. Every profile is a TOML file with args, env vars, and hardware metadata. The catalog matches them to your machine.
Profiles are starting points, not guarantees. Hardware differs — expect to adapt. But starting from someone's working config beats starting from an empty terminal.
Find a profile for your machine →26B-A4B: 16-18 GB (4-bit), 28-30 GB (8-bit), 52 GB (BF16/FP16).
26B-A4B: 16-18 GB (4-bit), 28-30 GB (8-bit), 52 GB (BF16/FP16).
26B-A4B: 16-18 GB (4-bit), 28-30 GB (8-bit), 52 GB (BF16/FP16).
26B-A4B: 16-18 GB (4-bit), 28-30 GB (8-bit), 52 GB (BF16/FP16).
E2B: 4 GB (4-bit) or 5-8 GB (8-bit), phone/edge.
E2B: 4 GB (4-bit) or 5-8 GB (8-bit), phone/edge.
Not a list of what fits. Not per-backend defaults. The exact config someone tuned for your model and hardware — imported and run with one command.
Profiles are TOML files with a documented schema. They live in GitHub, not a service. The catalog is a thin index over real source files.
An import is one shell command. The same TOML produces the same args and env on every machine — no hidden web of preferences.
Share the args your machine and model converged on. PR-only. No accounts.