Every open model in our directory, ranked by how well it runs on Apple M3 Ultra (32-core CPU, 80-core GPU) (512GB) at 4-bit. Grades weigh whether the model fits in memory and how fast it should run.
Apple M3 Ultra (32-core CPU, 80-core GPU) has 512 GB of VRAM, enough to run 166 of the 169 open models we track at 4-bit. The largest that still fits well is LongCat-2.0 (1600B), and the top-rated is Mixtral 8x7B Instruct at about 58 tok/s.
166 of 169 models run comfortably on Apple M3 Ultra (32-core CPU, 80-core GPU).
See full Apple M3 Ultra (32-core CPU, 80-core GPU) specs and pricingThe top models this device can run at 4-bit, ranked by fit and speed.
| Model | Grade | Speed | VRAM |
|---|---|---|---|
| Mixtral 8x7B InstructMistral AI | SS | 58.0 tok/s | 11.4 GB |
| DiffusionGemma 26B-A4BGoogle | AA | 62.9 tok/s | 10.5 GB |
| Gemma 4 26B-A4B ITGoogle | AA | 59.9 tok/s | 11.0 GB |
| North Mini CodeCohere | AA | 78.6 tok/s | 8.4 GB |
| Nemotron 3 Nano Omni |
Mixtral 8x7B Instruct is the highest-rated model that runs on Apple M3 Ultra (32-core CPU, 80-core GPU) at 4-bit, at about 58 tok/s. In total, Apple M3 Ultra (32-core CPU, 80-core GPU) can run 166 of the 169 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.
| AA |
| 77.3 tok/s |
| 8.5 GB |
| Qwen3.6 35B-A3BAlibaba | AA | 77.3 tok/s | 8.5 GB |
| Qwen3.5-35B-A3BAlibaba | AA | 77.3 tok/s | 8.5 GB |
| Qwen3-30B-A3BAlibaba | AA | 122.4 tok/s | 5.4 GB |
| Llama 2 13B ChatMeta | AA | 77.9 tok/s | 8.5 GB |
| Llama 3.1 8B InstructMeta | AA | 49.5 tok/s | 13.3 GB |
| Llama 3 8B InstructMeta | AA | 116.4 tok/s | 5.7 GB |
| Carnice-9b for Hermes agentkai-os | AA | 109.6 tok/s | 6.0 GB |
Mixtral 8x7B InstructMistral AI | 46.7B(12.9B active) | SS | 58.0 tok/s | 11.4 GB | |
DiffusionGemma 26B-A4BGoogle | 25.2B(3.8B active) | AA | 62.9 tok/s | 10.5 GB | |
Gemma 4 26B-A4B ITGoogle | 26B(4B active) | AA | 59.9 tok/s | 11.0 GB | |
North Mini CodeCohere | 30B(3B active) | AA | 78.6 tok/s | 8.4 GB | |
Nemotron 3 Nano OmniNVIDIA | 30B(3B active) | AA | 77.3 tok/s | 8.5 GB | |
Qwen3.6 35B-A3BAlibaba | 35B(3B active) | AA | 77.3 tok/s | 8.5 GB | |
Qwen3.5-35B-A3BAlibaba | 35B(3B active) | AA | 77.3 tok/s | 8.5 GB | |
Qwen3-30B-A3BAlibaba | 30B(3B active) | AA | 122.4 tok/s | 5.4 GB | |
Ad CodeRabbitAI-powered Code Reviews. Cut review time & bugs in half, instantly.Try for Free | |||||
Llama 2 13B ChatMeta | 13B | AA | 77.9 tok/s | 8.5 GB | |
| 8B | AA | 49.5 tok/s | 13.3 GB | ||
| 8B | AA | 116.4 tok/s | 5.7 GB | ||
| 9B | AA | 109.6 tok/s | 6.0 GB | ||
LFM2.5-8B-A1BLiquid AI | 8.3B(1.5B active) | AA | 226.9 tok/s | 2.9 GB | |
Gemma 4 E4B ITGoogle | 4B | AA | 95.3 tok/s | 6.9 GB | |
Gemma 3 4B ITGoogle | 4B | AA | 95.3 tok/s | 6.9 GB | |
PersonaPlex 7BNVIDIA | 7B | AA | 137.7 tok/s | 4.8 GB | |
Ad Vast.aiAffordable on-demand GPU rentals for training and inference. Pick from thousands of hosts.Rent a GPU | |||||
Llama 2 7B ChatMeta | 7B | AA | 137.7 tok/s | 4.8 GB | |
Mistral 7B InstructMistral AI | 7B | AA | 103.1 tok/s | 6.4 GB | |
minimax-m2.5MiniMax | 230B(10B active) | AA | 29.0 tok/s | 22.7 GB | |
Gemma 4 E2B ITGoogle | 2B | AA | 177.8 tok/s | 3.7 GB | |
VibeThinker-3BWeiboAI | 3B | AA | 172.9 tok/s | 3.8 GB | |
Qwen3.5-122B-A10BAlibaba | 122B(10B active) | AA | 24.2 tok/s | 27.3 GB | |
Falcon 40B InstructTechnology Innovation Institute | 40B | AA | 27.1 tok/s | 24.4 GB | |
Qwen3.5-9BAlibaba | 9B | AA | 26.8 tok/s | 24.6 GB | |
Ad RunPodServerless and dedicated GPU cloud built for AI workloads. Spin up instances in seconds.Launch on RunPod | |||||
Qwen3-235B-A22BAlibaba | 235B(22B active) | BB | 18.1 tok/s | 36.3 GB | |
Model Compatibility
Use the calculator to weigh this device against any other, model by model, with speed and fit scores.
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.
