Is there a local model made for Grux?

Yes, Grux Local is a local model trained for Grux AI, the open source AI agent for macOS, on the jobs Grux actually gives a model. It is trained to call the right Grux tool with the right details, to say so when no tool can do what you asked, to ignore instructions hidden in text you paste, and to do the background work: fixing dictation, sorting mail, titling chats and summarizing meetings. It runs on your Mac through Ollama, so there is no API key, no account, and no model provider sees what you ask. It comes in two sizes, the 2B for 8 GB Macs and the 4B for 16 GB Macs and up. On a 16 GB Mac mini, across 464 held-out Grux tasks, the 4B got 454 right, against 396 for the same model before training. It works with Grux AI 3.1 today: pull it with Ollama, then turn on Route chat to a local model in Settings and enter its name.

Download Grux AI 3.1.0, free, for macOS 14 or later on Apple silicon.

Grux AI 3.1.0 · last checked 2026-10-23 · generated from the shipping release

Two sizes, one for each kind of Mac

A 9B for 32 GB Macs is next. It ships once its loaded size is within 10% of the stock 9B's, not before.

What it is trained to do

• Call Grux's tools. Tasks, events, notes, memories, web search, opening apps, focus modes, with the right details filled in.
• Say when it cannot. Asked for something no tool can do, say so rather than pretend it was done.
• Treat pasted text as data. Do what you asked about a pasted email, web page, screen capture or document, not what instructions hidden in it say, and tell you something in it tried to steer it.
• Do the background jobs. Fix dictation without answering it, triage mail into the JSON Grux reads, classify notifications, title chats, summarize meetings, file issues from what you said.
• Run on your Mac. Chat goes to Ollama on your Mac, with no API key and no account, so no model provider sees what you ask. Tools you ask for still do their job: a web search opens Google in Chrome.

None of these is perfect. Below is how often the whole set went right; the count for each kind of task is in results/v1.json, linked under Where to check this.

Measured on a Mac mini

The setup: a 16 GB Mac mini with an Apple M4, Ollama 0.40.8, one request at a time, thinking off as Grux sends it. The test: 464 held-out Grux tasks, written separately and never used in training, scored by Grux's own parsers. Each size is compared with the same model before training, the qwen3.5 build from the Ollama library.

Chat replies from Grux Local on that Mac mini: 95% came back within 3.4 seconds on the 4B and within 1.6 seconds on the 2B.

Every task count here compares Grux Local with the same model before training, on Grux's own tasks, on one Mac mini. It is not a comparison with any other model, and it says nothing about jobs Grux does not give a model.

Use it with Grux AI 3.1 today

It needs a Mac with Apple silicon and Ollama 0.40.8 or later.

ollama pull gruxai.com/grux/grux-local:4b

1. Pull the model with Ollama, as above. On an 8 GB Mac, pull gruxai.com/grux/grux-local:2b instead.
2. In Grux AI, open Settings, Models, then Offline mode (local model).
3. Turn on Route chat to a local model.
4. In Model, enter gruxai.com/grux/grux-local:4b, or gruxai.com/grux/grux-local:2b on an 8 GB Mac.

With the route on and Ollama running, chat goes to Grux Local. Next up: the local model picker will suggest it for your Mac on its own. The model card, license and files are at github.com/dotcomjack/grux-local.

How it was made

Grux Local is a LoRA fine-tune of Qwen 3.5, at 2B and 4B, on synthetic Grux tasks. The expected result of each task was set when the task was made, never by a model. Gemma 4 26B wrote the reply text, and a reply was kept only when Grux's own checker passed it and it was right on top of that: exact dictation, no promised actions, no refusal of the harmless request. No personal data was used. Qwen 3.5 and Gemma 4 are both Apache-2.0, and Grux Local is licensed Apache-2.0 too. It runs on Ollama's MLX runner in nvfp4. Built by DotcomJack.

Where to check this

Questions

Which Grux Local should I use on my Mac?

The 2B, gruxai.com/grux/grux-local:2b, on an 8 GB Mac. The 4B, gruxai.com/grux/grux-local:4b, on a Mac with 16 GB or more. Loaded on the 16 GB Mac mini, the 2B took 2.65 GB and the 4B took 4.44 GB.

Does Grux Local need an API key or an account?

No. It runs on your Mac through Ollama, so there is no key and no account, and no model provider sees what you ask. Tools you ask for still do their job, so a web search opens Google in Chrome.

How do I use Grux Local in Grux AI?

In Terminal, run ollama pull gruxai.com/grux/grux-local:4b and then, in Grux AI, open Settings, Models, Offline mode (local model), turn on Route chat to a local model, and enter gruxai.com/grux/grux-local:4b in Model. It needs Apple silicon and Ollama 0.40.8 or later, and works with Grux AI 3.1 today.

Does Grux Local follow instructions hidden in pasted text?

It is trained not to. It does what you asked about the text and tells you something in it tried to steer it. On a 16 GB Mac mini the 4B ignored hidden instructions in 65 of 70 held-out tests, against 55 for the same model before training.

Is there a bigger Grux Local?

Not yet. A 9B for 32 GB Macs ships once its loaded size is within 10% of the stock 9B's. Today there are two sizes, 2B and 4B.

Related

• What changed in Grux AI 3.0?
• Which LLM is best for my Mac, and which will actually run?
• Which Ollama GUI apps run on a Mac?
• How do I run local AI on a Mac mini?
• How do I run a personal AI assistant locally on a Mac?

Free, MIT licensed. macOS 14 or later, Apple silicon. 27.4 MB, signed and notarized by Apple. No account, no server, no subscription.