Which AI models can the Apple M3 Ultra (256GB) run?
Every model we track, checked against the Apple M3 Ultra (256GB) at 8K context. The largest that runs well is DeepSeek V4 Flash 284B-A13B.
Unified memory256 GB192 GB usable by the GPU
Bandwidth819GB/s
Runs well60/69models at Q4_K_M or better
Fastest model~328tok/s · Qwen3 0.6B
Every model on the Apple M3 Ultra (256GB)
Bars show memory used out of 192 GB. Pick a row for the full breakdown.
60 Runs well
NameMemory usedBest quantSpeed
DeepSeek V4 Flash 284B-A13B284BQ4_K_M · 173 GB~62 tok/s
Qwen3 235B-A22B235BQ5_K_M · 169.4 GB~31 tok/s- Qwen3.8 Flash Next180BQ6_K · 148.4 GB~100 tok/s
gpt-oss 120B117BQ8_0 · 125 GB~91 tok/s
Llama 4 Scout 17B-16E109BQ8_0 · 118 GB~27 tok/s
GLM-4.5-Air 106B-A12B106BQ8_0 · 114.8 GB~39 tok/s- Qwen3 Coder Next 80B-A3B80BFP16 · 160.2 GB~82 tok/s
- Qwen2.5 72B73BFP16 · 148.7 GB~3 tok/s
- Llama 3.1 70B71BFP16 · 144.5 GB~3 tok/s
- Llama 3.3 70B71BFP16 · 144.5 GB~3 tok/s
- DeepSeek R1 Distill Llama 70B71BFP16 · 144.5 GB~3 tok/s
Mixtral 8x7B47BFP16 · 95.1 GB~19 tok/s- Qwen3.6 35B-A3B36BFP16 · 72.8 GB~82 tok/s
- Qwen2.5 32B33BFP16 · 68.3 GB~7 tok/s
- DeepSeek R1 Distill Qwen 32B33BFP16 · 68.3 GB~7 tok/s
- Qwen3 32B33BFP16 · 68.3 GB~7 tok/s
- QwQ 32B33BFP16 · 68.3 GB~7 tok/s
- Qwen2.5 Coder 32B33BFP16 · 68.3 GB~7 tok/s
Nemotron 3 Nano 30B-A3B32BFP16 · 63.9 GB~70 tok/s
Gemma 4 31B31BFP16 · 64.4 GB~8 tok/s- GLM-4.7 Flash 30B-A3B31BFP16 · 63.4 GB~82 tok/s
- Qwen3 30B-A3B31BFP16 · 62.4 GB~74 tok/s
- Qwen3 Coder 30B-A3B31BFP16 · 62.4 GB~74 tok/s
- Qwen3.8 27B28BFP16 · 56.7 GB~9 tok/s
- Qwen3.6 27B28BFP16 · 56.7 GB~9 tok/s
- Gemma 3 27B27BFP16 · 56.5 GB~9 tok/s
- Gemma 2 27B27BFP16 · 57.3 GB~9 tok/s
- Gemma 4 26B-A4B26BFP16 · 52.5 GB~65 tok/s
- Mistral Small 3.2 24B24BFP16 · 49.9 GB~10 tok/s
- Mistral Small 24B24BFP16 · 49.1 GB~10 tok/s
- Devstral Small 24B24BFP16 · 49.1 GB~10 tok/s
- Magistral Small 24B24BFP16 · 49.1 GB~10 tok/s
- gpt-oss 20B21BFP16 · 42.6 GB~68 tok/s
- Qwen3 14B15BFP16 · 31.5 GB~17 tok/s
- Qwen2.5 Coder 14B15BFP16 · 31.8 GB~17 tok/s
- DeepSeek R1 Distill Qwen 14B15BFP16 · 31.8 GB~17 tok/s
- Qwen2.5 14B15BFP16 · 31.6 GB~17 tok/s
Phi-4 14B15BFP16 · 31.7 GB~17 tok/s- Phi-4 Reasoning 14B15BFP16 · 31.7 GB~17 tok/s
- Mistral Nemo 12B12BFP16 · 26.3 GB~20 tok/s
- Gemma 3 12B12BFP16 · 25.9 GB~20 tok/s
- Gemma 4 12B12BFP16 · 24.9 GB~21 tok/s
- Gemma 2 9B9.2BFP16 · 21.2 GB~27 tok/s
- Qwen3 8B8.2BFP16 · 18.2 GB~30 tok/s
- DeepSeek R1 0528 Qwen3 8B8.2BFP16 · 18.2 GB~30 tok/s
Granite 3.3 8B8.2BFP16 · 18.3 GB~30 tok/s- Llama 3.1 8B8BFP16 · 17.7 GB~31 tok/s
- Qwen2.5 7B7.6BFP16 · 16.3 GB~32 tok/s
- Qwen2.5 Coder 7B7.6BFP16 · 16.3 GB~32 tok/s
- DeepSeek R1 Distill Qwen 7B7.6BFP16 · 16.3 GB~32 tok/s
- Mistral 7B v0.37.3BFP16 · 16.2 GB~34 tok/s
- Gemma 3 4B4.3BFP16 · 9.5 GB~57 tok/s
- Qwen3 4B4BFP16 · 9.8 GB~61 tok/s
- Phi-4 Mini 3.8B3.8BFP16 · 9.4 GB~64 tok/s
- Phi-3.5 Mini 3.8B3.8BFP16 · 11.5 GB~64 tok/s
- Llama 3.2 3B3.2BFP16 · 8 GB~77 tok/s
- Qwen3 1.7B2BFP16 · 5.6 GB~121 tok/s
- Llama 3.2 1B1.2BFP16 · 3.3 GB~198 tok/s
- Gemma 3 1B1BFP16 · 2.6 GB~246 tok/s
- Qwen3 0.6B0.8BFP16 · 3 GB~328 tok/s
2 Fits with heavy quantization
NameMemory usedBest quantSpeed
7 Too big
NameMemory usedBest quantSpeed
Kimi K3.1 Pro 2.8T-A104B2800Bneeds 1698.3 GB192 GB free- Qwen3.8 2.4T-A95B2446Bneeds 1484.3 GB192 GB free
- DeepSeek V4 Pro 1.6T-A49B1600Bneeds 971.2 GB192 GB free
- Mistral Large 4 1T-A52B1000Bneeds 607.5 GB192 GB free
- DeepSeek V4.1 Flash763Bneeds 463.4 GB192 GB free
- GLM-5.3753Bneeds 457.8 GB192 GB free
- GLM-5 744B-A40B744Bneeds 452.4 GB192 GB free
Run it in the cloud
Need more than 256 GB? Rent a bigger GPU by the hour.
Rent a cloud GPU by the hour. No upfront cost, stop any time.