Models

Open models you can actually run

Parameter counts, context length, licence and measured VRAM use, plus the one command that starts the model locally. Anything we have not run ourselves is labelled as vendor-reported.

1 pagesUpdated 2026-10-23

Nimbus 7B

A 7.2B instruction model with an Apache 2.0 licence and an 8K context window. Small enough to run on a single consumer card, and the first open model in a while whose quantised output does not fall apart on structured tasks.

Sample
Text models
Local deployment

How this section works

Model pages are written to answer one question before you download anything: will this run on the hardware I already own. That is why the VRAM table is the first thing on the page and why every row names the card it was measured on. Vendor-reported figures are kept in a separate table and labelled, because a parameter count tells you almost nothing about whether the model fits.

Each page also carries the command that actually starts the model, in the two runners most people use, with the flag that trips everyone up noted underneath. If a model needs a specific quantisation to fit in 8GB, that is said plainly rather than left in a comment. Pages get revised when a new quantisation or a breaking change in the runner lands.