Every open model in our directory, ranked by how well it runs on Mac Studio with M5 Ultra (512GB) at 4-bit. Grades weigh whether the model fits in memory and how fast it should run.
Mac Studio with M5 Ultra has 512 GB of VRAM, enough to run 385 of the 419 open models we track at 4-bit. The largest that still fits well is Mistral Large 3 675B (675B), and the top-rated is NorthCUA-Pro 397B at about 88 tok/s.
385 of 419 models run comfortably on Mac Studio with M5 Ultra.
See full Mac Studio with M5 Ultra specs and pricingThe top models this device can run at 4-bit, ranked by fit and speed.
| Model | Grade | Speed | VRAM |
|---|---|---|---|
| NorthCUA-Pro 397BNorthCUA | SS | 88.5 tok/s | 231.3 GB |
| K2 Horizon 375B-A23BIFM | SS | 66.2 tok/s | 218.8 GB |
| MemT-1.0-FlashMemTensor | SS | 88.5 tok/s | 168.1 GB |
| minimax-m2.5MiniMax | SS | 42.6 tok/s | 150.3 GB |
| AliceAI-Foundation-80B-A3B-Baseyandex | SS | 113.2 tok/s | 53.9 GB |
| Nemotron 3.5 Bolt 74B-A9B 24GBNVIDIA | SS | 160.6 tok/s | 43.7 GB |
| NVIDIA Nemotron 3.5 Bolt 74B-A9B 24GBNVIDIA | SS | 160.6 tok/s | 43.7 GB |
| Comet 60BMultiverse Computing | SS | 115.8 tok/s | 36.4 GB |
| Mixtral 8x7B InstructMistral AI | SS | 85.0 tok/s | 31.0 GB |
| KolibriAleph Alpha | SS | 36.8 tok/s | 69.7 GB |
| K2-Horizon-MoVA-36B-A4BInstitute of Foundation Models | SS | 49.3 tok/s | 38.2 GB |
| Qwen3.5-122B-A10BAlibaba | SS | 35.4 tok/s | 92.2 GB |
NorthCUA-Pro 397BNorthCUA | 397B(17B active) | SS | 88.5 tok/s | 231.3 GB | |
| 375B(23B active) | SS | 66.2 tok/s | 218.8 GB | ||
MemT-1.0-FlashMemTensor | 288B(17B active) | SS | 88.5 tok/s | 168.1 GB | |
minimax-m2.5MiniMax | 230B(10B active) | SS | 42.6 tok/s | 150.3 GB | |
| 81.29B(3B active) | SS | 113.2 tok/s | 53.9 GB | ||
| 74B(9B active) | SS | 160.6 tok/s | 43.7 GB | ||
| 74B(9B active) | SS | 160.6 tok/s | 43.7 GB | ||
Comet 60BMultiverse Computing | 61.1B(12.8B active) | SS | 115.8 tok/s | 36.4 GB | |
| Ad | |||||
Mixtral 8x7B InstructMistral AI | 46.7B(12.9B active) | SS | 85.0 tok/s | 31.0 GB | |
KolibriAleph Alpha | 78B(3B active) | SS | 36.8 tok/s | 69.7 GB | |
K2-Horizon-MoVA-36B-A4BInstitute of Foundation Models | 36B(4B active) | SS | 49.3 tok/s | 38.2 GB | |
Qwen3.5-122B-A10BAlibaba | 122B(10B active) | SS | 35.4 tok/s | 92.2 GB | |
Recursive FlintRecursive | 39.8B(2.5B active) | SS | 134.3 tok/s | 28.8 GB | |
Qwen3.6 35B-A3BAlibaba | 35B(3B active) | SS | 113.2 tok/s | 27.1 GB | |
Qwen3.5-35B-A3BAlibaba | 35B(3B active) | SS | 113.2 tok/s | 27.1 GB | |
MiniCPM-V 4.7 35B-A3Bopenbmb | 35.21B(3B active) | SS | 113.2 tok/s | 27.2 GB | |
| Ad | |||||
Raon-1 mini previewKRAFTON | 28.78B(4.3B active) | SS | 80.4 tok/s | 26.2 GB | |
Xing4.0-29B-A4BChina Telecom AI | 29B(4B active) | SS | 86.2 tok/s | 25.7 GB | |
DiffusionGemma 26B-A4BGoogle | 25.2B(3.8B active) | SS | 92.1 tok/s | 22.9 GB | |
Gemma 4 26B-A4B ITGoogle | 26B(4B active) | SS | 87.7 tok/s | 23.8 GB | |
North Mini CodeCohere | 30B(3B active) | SS | 115.2 tok/s | 24.0 GB | |
Nemotron 3 Nano OmniNVIDIA | 30B(3B active) | SS | 113.2 tok/s | 24.2 GB | |
Qwen3-30B-A3BAlibaba | 30B(3B active) | SS | 179.4 tok/s | 21.0 GB | |
Holo4-35B-A3BHcompany | 35B(3B active) | SS | 413.1 tok/s | 20.9 GB | |
| Ad | |||||
LFM2.5-24B-A2BLiquid AI | 24B(2.45B active) | AA | 215.1 tok/s | 17.0 GB | |

Model Compatibility
Use the calculator to weigh this device against any other, model by model, with speed and fit scores.
Straight answers to the questions we hear most often.
Can't find what you're looking for? Book a Discovery Call
NorthCUA-Pro 397B is the highest-rated model that runs on Mac Studio with M5 Ultra at 4-bit, at about 88 tok/s. In total, Mac Studio with M5 Ultra can run 385 of the 419 open models we track.
We size each model at 4-bit (Q4_K_M), add the KV cache for its context length and a fixed runtime overhead, then compare that to this device VRAM. Each model is graded on how comfortably it fits and how fast it should run on this card.
It is an estimate of decode speed in tokens per second at 4-bit, based on this device memory bandwidth and each model size. Real speed depends on your runtime, batch size, and context length, so use it to compare models, not as a hard promise.
Lower-bit quantization shrinks a model so more of them fit. These grades already assume 4-bit, which is the common local default. Going below 4-bit can squeeze a larger model in at some quality cost; going above needs more VRAM and may not fit.
Mixture-of-experts models only activate a fraction of their parameters at a time, so their runtime memory is closer to the active size than the headline size. We score fit on the active parameters, which is why some very large models still fit.