The AIC Lab
Memory decides which models fit. This is what we run on each size of Mac, and what we measured in our lab.
On Apple Silicon, the CPU, GPU and Neural Engine share one pool of memory, and a model has to fit in it. Each size below runs everything the sizes before it run.
16 GB
MacBook Pro 16-inch with M1 Pro
Our open-weight model list starts at 18 GB, so a 16 GB Mac stays with Apple's on-device model. It can also link to a larger AI server on your network.
64 GB
Mac mini with M5 Pro
Room for Qwen 3.8, Gemma 4 and the smaller Q4 builds of Muse Glimmer. The full BF16 build of Muse Glimmer needs 96 GB.
See it at AIC Hardware →Or get it set up as our Compact server →
128 GB
MacBook Pro 14-inch with M5 Max
Everything the 64 GB tier runs, plus the full BF16 builds of Qwen 3.8 and Muse Glimmer. Most of the numbers on this page were measured on this MacBook.
512 GB
Mac Studio with M3 Ultra
Adds Qwen 3.8 Flash-Next and GLM 5.3 Flash, which are too large for our 128 GB MacBook.
See it at AIC Hardware →Or get it set up as our Studio server →
Two Macs, 1 TB
Two Mac Studios with M3 Ultra
GLM-5.3 and DeepSeek V4.1 Flash are too big for one Mac, so they run split across both over Thunderbolt 5. They come fully assembled in one rack.
Our own runs, on our own Macs, in October 2026. Rounded, and only for the Macs named.
| Model | Build | Peak memory | Smallest Mac | Speed | Tested on |
|---|---|---|---|---|---|
| Muse Glimmer 30B | Q4_K_M | 22 GB | 32 GB | 23 tokens/s | MacBook Pro, M5 Max, 128 GB |
| Muse Glimmer 30B | Q4_K_XL | 25 GB | 32 GB | 19 tokens/s | MacBook Pro, M5 Max, 128 GB |
| Gemma 4 12B | BF16 | 27 GB | 32 GB | 19 tokens/s | MacBook Pro, M5 Max, 128 GB |
| Qwen 3.8 27B | FP8 | 46 GB | 64 GB | 17 tokens/s | MacBook Pro, M5 Max, 128 GB |
| Muse Glimmer 30B | BF16 | 61 GB | 96 GB | 12 tokens/s | MacBook Pro, M5 Max, 128 GB |
| Qwen 3.8 27B | BF16 | 68 GB | 96 GB | 14 tokens/s | MacBook Pro, M5 Max, 128 GB |
| Qwen 3.8 Flash-Next | FP8 | 193 GB | 256 GB | 26 tokens/s | Mac Studio, M3 Ultra, 512 GB |
| GLM 5.3 Flash | FP8 | 352 GB | 512 GB | 20 tokens/s | Mac Studio, M3 Ultra, 512 GB |
| GLM-5.3 | FP8 | 463 GB per Mac | Two 512 GB Macs | 20 tokens/s* | Two Mac Studios, M3 Ultra, 512 GB each |
| DeepSeek V4.1 Flash | MXFP4 | 285 GB per Mac | Two 512 GB Macs | 34 tokens/s* | Two Mac Studios, M3 Ultra, 512 GB each |
Speed depends on the Mac. A Mac with less memory bandwidth, like the Mac mini, will be slower than the Macs listed here. *Split models show the median across 22 real Izuran Pro prompts, and the peak on one Mac.
These are the checks a model goes through before it enters our catalog.
Step 1
Can this Mac hold it?
We send a request of about 120,000 tokens and record the peak memory. That sets the smallest Mac we list a model for.
Step 2
Can it use tools?
We run it through Izuran's own tool tasks, using the model's own template, and through longer agent tasks.
Step 3
Can it read long documents?
We hide facts inside long documents and check that it finds them on the first try.
Step 4
Is it fast enough to use?
We time how fast it writes with 8,000 tokens of context, and how long it takes to read 60,000 tokens.
In private testing
Our private AI assistant for Mac, iPhone and Apple Watch. The models run on your own Macs.
Join the waitlist →Set up for your office
A server on your network with Izuran Pro and your models installed, for your whole team.
See the server →Buy the hardware
Used Macs from the AIC Lab, tested, erased and graded.
Visit the store →