runlocal-sh
v0.1.0
Published
Pick the right local AI model + runtime for your hardware, in one command. Will it run? Ask runlocal.
Maintainers
Readme
runlocal
Pick the right local AI model for your hardware. Will it run? Ask runlocal.
npx runlocal-sh # detect your rig → models that fit
npx runlocal-sh can-i-run deepseek-r1 # will a specific model run?
npm install -g runlocal-sh # or install globally$ runlocal can-i-run deepseek-r1
✅ Runs (tight) DeepSeek-R1 (distill) 32B @ Q3_K_M
…needs ~17.6 GB of your 18.4 GB usable unified memory. Expect ~6–10 tok/s.
→ runlocal install deepseek-r1:32bCommands
runlocal— fingerprint your machine + recommend models that actually fitrunlocal can-i-run <model>— verdict for a model: best quant + tok/s rangerunlocal install <model>— install the best-fitting quant via Ollamarunlocal bench <model>— re-measure an installed model: real tok/s vs the predicted rangerunlocal list— catalog models annotated for your rigrunlocal graph [model]— knowledge graph of the catalog →graph.json+GRAPH_REPORT.md+ interactivegraph.htmlrunlocal doctor— diagnose your setup (GPU, runtimes, fixes)runlocal advise <models>— which rigs run ALL the models a workflow needs (comma-separated), smallest memory firstrunlocal launch-post <model>— markdown launch post: engine-computed verdict table across popular rigs + links
Flags
--for <use-case> · --context <n> · --hardware <id> (simulate a rig) ·
--depth <n> / --out <dir> / --query "<q>" (graph) · --json · --no-telemetry ·
--no-measure (skip the post-install measured-vs-predicted accuracy check) · -y/--yes · --all
How verdicts are computed
Real fit math, not a hand-wave: GQA-aware KV-cache sizing, calibrated bits-per-weight per quant, and a conservative usable-memory budget (Apple's RAM-size-dependent Metal wired limit; NVIDIA VRAM headroom). Speed is a memory-bandwidth-derived range, never a fake exact number.
Apple Silicon is excellent in v0.1; NVIDIA is best-effort; other setups fall back to CPU estimates.
Privacy
Collects nothing unless you opt in (then: anonymous rig + model choice, plus a
measured-vs-predicted speed datapoint after an install or a bench run — no PII).
Disable with --no-telemetry, RUNLOCAL_TELEMETRY=0, or DO_NOT_TRACK=1.
Source + web checker: https://runlocal.sh · MIT © Raybeam
