local-rank

About local-rank

local-rank tracks the open-source runtimes that serve large language models on your own hardware — laptops, desktops, servers, even phones. We focus on the engine layer, the software that takes a quantized model file and turns it into tokens.

The leaderboard is not a benchmark. It is a community-velocity signal: which runtimes are gaining stars, shipping releases, and merging commits this week.

The catalog is hand-curated and intentionally narrow. If a project you care about is missing, file an issue on the upstream repository referencing this page — we add new runtimes as the ecosystem grows.

What we do not do

• We do not benchmark tokens/second.
• We do not maintain a model catalog.
• We do not rank hosted inference providers.

Methodology

Every metric on this page is fetched from the GitHub REST API at runtime. No values are hardcoded, mocked, or fabricated.

Refresh cadence

Hourly when a GitHub token is configured (the production deployment). Every four hours otherwise.

Star deltas

The Δ7d and Δ30d columns compare today's star count to the oldest snapshot inside the window. If the runtime has been tracked for less than the window length, the delta is left blank.

Velocity score

A composite 0–100 score combining relative star momentum, release cadence, commit cadence, contributor breadth, and the recency of the latest release and commit. See lib/scoring.js in the source for exact weights.

Sources

• GET /repos/{owner}/{repo}
• GET /repos/{owner}/{repo}/releases?per_page=100
• GET /repos/{owner}/{repo}/commits?per_page=1&since={iso}
• GET /repos/{owner}/{repo}/contributors?per_page=1&anon=true