Documentation
lean is a single-binary inference runtime for very large Mixture-of-Experts
models. It streams expert weights on demand across VRAM, RAM, and NVMe, so a
122B-parameter model runs on one 24 GB consumer GPU instead of a rack.
Models ship in .lmpack format: a small always-resident core plus one file per
expert. The runtime keeps the hot experts in VRAM, the warm ones in page cache,
and pages the rest in from disk as routing demands them.
Start here
Section titled “Start here”- Quick Start - install, pull a model, and get a first response.
- CLI reference - every
leansubcommand and flag. - OpenAI-compatible API -
lean serveendpoints, request and response shapes, and what is not supported yet. - Aider - drive a local model from a real coding agent.
What you need
Section titled “What you need”| Model | Total params | Minimum VRAM | Tier |
|---|---|---|---|
| lean-agent-light | 35B | 12 GB | Free |
| lean-coder-welter | 80B | 12 GB | Free |
| lean-agent-middle | 122B | 24 GB | Paid |
More RAM buys you throughput rather than capability: experts that fit in page cache never touch the disk. See Models for the full catalog and Benchmarks for measured throughput on specific hardware.
Not sure what your machine can take? lean tune reads your GPUs, RAM, and free
disk, then recommends a model and settings.