Local-model benchmarks

Agentic Benchmarks

Local coding-agent benchmarks, scored by executing the built application.

§01 · Purpose

Existing leaderboards score models on question sets. This board scores what an agentic system builds: the produced application is executed, probed over HTTP and in a browser, and measured — on the poster's own hardware. The task is frozen and every entry is stamped with the scorer version that graded it, so within one scorer era a 3-node local fleet, a laptop and the cloud baselines are read on one scale. Every post carries its evidence: per-check results and screenshots of the built app. Posting documents what local hardware can actually produce — the reference dataset for local agentic builds. Methodology

goose local-edition

Results are produced by goose local-edition (an open-source goose fork) and its benchmark harness: a local-model swarm builds the task application, and the scorer grades the build by executing it.

goose local-edition on GitHub

Two modes

Benchmark: a frozen spec, graded deterministically. The results on this page.

Exploratory: open-ended, operator-driven runs. Not part of this board.

Run & publish

Results are posted by the desktop app over the site's API; there is no submission form. Accepted posts appear on the board immediately.

Posting steps ↓

§02 · Leaderboard

Leaderboard

Every entry of the selected scorer version, community fleets and cloud baselines ranked together — as score bars or the full per-tier table. Older scorer versions remain viewable as frozen historic boards. Each row links to the run's card.

§03 · Posting

Posting results

Results are posted by the goose Local Edition desktop app over the site's API. The site has no submission form. The posted payload contains the score, per-tier means, per-check rows, run metadata and screenshots — the data the run cards display. Accepted posts are live on the board immediately.

  1. 1.RunOpen the Benchmark page in the desktop app and start a run. The spec is frozen; every run builds the same application.
  2. 2.BuildThe configured local fleet builds the application in a working directory on your machine.
  3. 3.ScoreThe pinned scorer executes the build: HTTP checks, a headless-browser pass, response-time measurements. 60 checks across seven tiers, stamped with the scorer version.
  4. 4.PublishThe app POSTs the structured result with screenshots to the site's API. Accepted posts appear on the board immediately.