Grux AI Download

Which Ollama model should I pull on a Mac?

Pull the largest model your Mac has memory for, then check it can do the job you need done. On Apple silicon that usually means a 4B build on a 16 GB Mac and a 9B from 24 GB. The Ollama library tells you how big a download is; it does not tell you whether a model calls the right tool with the right details, says so when no tool fits, or notices when pasted text tries to steer it. Grux scores small open models from that library on its own held-out set of 464 tasks, written separately and never used in training, and publishes each count with a 95% interval and the raw JSON behind it. The tables below join the two halves: what a pull costs, and how many of those tasks it got right, with the date, the machine and the runner beside it. 3 library builds are scored on the weekly run, from 3.3 GB to 6.6 GB.

Download Grux AI 3.3.1, free, for macOS 14 or later on Apple silicon.

Grux AI 3.3.1 · last checked 2026-11-13 · generated from the shipping release

The library builds Grux has scored

Each row is the stock build pulled from the Ollama library with ollama pull, by the machine that scored it, on held-out set v1. The size on disk is what the build itself reports. Rows are ordered by tasks right, and the date links the raw result file it came from.

ModelSizeDiskLicenceRight of 464HiddenRaw
qwen3.5:9b9.0B, Q4_K_M6.6 GBApache-2.0396 of 46457 of 702026-11-10
qwen3.5:4b4.2B, Q4_K_M3.3 GBApache-2.0392 of 46451 of 702026-11-10
mimo-v2.6:9b9.0B, Q4_K_M6.6 GBMIT367 of 46459 of 702026-11-10

Runner for every row in this table: Ollama 0.40.16, llamacpp runner, thinking off, a 16,384 token context, one request at a time, on the one NVIDIA H200 GPU in the cloud, rented by the hour.

Every build in this table carries the same context length, 262,144 tokens, so it is not a column here.

The same tags, measured on a Mac

These are the same library tags pulled and scored on Apple silicon, on held-out set v1. They are kept in their own table because another machine and runner means another build of the weights, and two machines never share a table here. The counts differ from the table above, which is the point of keeping them apart.

ModelMachineRight of 464HiddenMeasuredRaw
qwen3.5:2bMac mini, Apple M4, 16 GB295 of 46443 of 702026-10-23JSON
qwen3.5:4bMac mini, Apple M4, 16 GB396 of 46455 of 702026-10-22JSON

Ollama 0.40.8, MLX runner, one request at a time, thinking off as Grux sends it

ModelMachineRight of 464HiddenMeasuredRaw
qwen3.5:9bMacBook Pro, Apple M4 Pro, 48 GB400 of 46465 of 702026-10-23JSON

Ollama 0.40.8, MLX runner, one request at a time, thinking off as Grux sends it

The runner changes the answers

A pull is weights plus whatever loads them, and the loader is part of the result, so every count on this page names the runner that produced it.

The Grux Local 4B is the one model here scored on the same Mac mini, Apple M4, 16 GB through more than one runner.

RunnerModelMachineRight of 464HiddenMeasuredRaw
Ollama 0.40.8, MLX runnergrux-local:4bMac mini, Apple M4, 16 GB454 of 46465 of 702026-10-23JSON
mlx-lm 0.32.2grux-local:4b16 GB Mac mini (Apple M4)458 of 46469 of 702026-11-15not published, see the hash below

The same model, packaged for each runner, on the same Mac and the same 464 tasks: 4 tasks and 4 hidden-instruction cases apart, one run each. The run behind the second row hashes to 9edfdfbe03da34a0. That is why every number on this page names its runner, and why a score read anywhere else is only about the runner it was measured on.

A score belongs to a runner, not to a model name. Ask which runner a number came from before you compare it with another one.

Picking by memory, before any score

What these counts do not say

Where to check this

ClaimSource
Every score in the library and Mac tables is a count on Grux's held-out set, with its raw JSON; the second runner's row names the hash of its run insteadGrux Local scoreboard
Both held-out sets are pinned by hash and size at the commit the rows came fromrecipe/check_prompts.py
The Mac rows and their task families are in the published result fileresults/v1.json
The library itself, with every model and its download sizeollama.com/library

Questions

Which Ollama model got the most of Grux's jobs right?
On the weekly table, qwen3.5:9b got 396 of 464. That is a count of Grux's own jobs on one machine and one runner, not a ranking of models and not a measure of general ability; the Grux Local scoreboard carries the interval around each weekly count, and two intervals that overlap may still differ.
How much disk does an Ollama model take?
The pull size is in the table above, read from the build that was scored. The smallest build scored here is 3.3 GB and the largest is 6.6 GB. A model also needs memory to work in once it is loaded, which is more than its file.
Does it matter how I run the model?
Yes. The same model, packaged for each runner, did not score the same on the same Mac, so every number on this page names the runner it was measured on, and the table above shows both rows.
Can I check these numbers myself?
Yes. Every row in the library and Mac tables links its raw JSON in the public grux-local repository at the commit this page was built from, and both held-out sets are pinned by hash there; the second runner's row names the hash of its run instead.
Download Grux AI 3.3.1 Read the source

Free, MIT licensed. macOS 14 or later, Apple silicon. 29.1 MB, signed and notarized by Apple. No account, no server, no subscription.