November 22, 2026

Omarchy gives local AI models their own panel with a memory bar

0

A new sparkles icon in the top bar opens a panel built around one question for local models: what is in memory, and will the next one fit?

The Omarchy agents panel before the change, listing Claude, Codex and Antigravity usage limits followed by local models such as Qwen, Z-Image Turbo and FLUX.2 Klein with their RAM needs

Image: dhh, omacom/omarchy PR #19597, before the change (GitHub)

Local AI models are moving out of the agents panel and into a panel of their own. DHH merged pull request #19597 into Omarchy’s quattro branch on November 21, 2026, and it’s coming in the next version. As the PR puts it, the agents panel “is about subscriptions and their limits, while a local model is about the hardware it runs on.”

The new bar icon, omarchy.local-models, shows sparkles and sits to the left of the agents icon. You can also open it with Super + Ctrl + Alt + X, the Alt variant of the agents panel’s Super + Ctrl + X, which a follow-up PR settled on. The icon only appears on a machine that can run a model itself: a recipe fits it, a model is already set up, or another Spark is cabled to it.

The centerpiece is a memory bar that splits the machine’s RAM, to scale, between Omarchy (the system and your apps), each running model, and what’s free. Free counts MemAvailable plus the GPU driver’s pool, the same room the engine uses to decide whether a model fits. Point at a running model and its slice lights up. Point at a stopped one and the bar lays what it needs into the free space, in the accent color when it fits and in the warning color when it doesn’t, with a label like “32 GB short”.

Each model gets a line that can start, stop, update, share and pair it, or open it in an agent or in Fantasi, with download and load progress shown in place. Lines can be dragged by their mark or moved with Ctrl + Up/Down, and the order sticks. The + button lists the catalog recipes that fit the machine alongside the tailnet’s hidden models, and only shows when there is something to add.

The command moves too: omarchy agent local … becomes omarchy models …. A new omarchy models use best <text|image|video|music> sets up the best model of that kind for the machine, and omarchy models use best on its own lists each kind’s pick. The PR notes that “Nothing local-model related shipped in v4.5.0,” so the old names get no compatibility shims.

The engine side landed the same day in omarchy-local-ai #84, which reports update progress, measures what each running model actually holds in memory rather than what it took to load, and recommends a best model of each kind.

The updated AI manual gives a sense of the numbers the bar will be juggling: Flash Next uses about 100 GiB with its full 256K context, qwen3.8-27b-gufo takes about 40 GiB on 128 GB Strix Halo machines, and qwen3.6-35b-a3b-llamacpp runs in about 30 GiB on a 64 GB laptop with Intel’s Arc B390. Omarchy runs every model with no network access.

It’s the latest step in Omarchy’s local AI push, after the curated recipe catalog, the TensorFold recipes for NVIDIA machines and model sharing across your tailnet.

Source: https://github.com/omacom/omarchy/pull/19597

Leave a Reply

Your email address will not be published. Required fields are marked *