All hardware / M3 Ultra 96GB
With 96 GB unified memory, the M3 Ultra 96GB runs 17 of 22 popular open models fully in fast memory at Q4_K_M or better (8K context). The largest that fit are Llama 4 Scout (Q4_K_M), Llama 3.3 70B (Q6_K), Hermes 4.3 36B (Q8_0). Fastest: Gemma 4 E2B at over 150 tokens/sec and Qwen3.6 35B-A3B at over 150 tokens/sec.
Check with your own RAM and context length| Model | Result | Best quant | Memory needed | Est. speed |
|---|---|---|---|---|
| Gemma 4 E2B | Runs on GPU | Q8_0 | 6.1 GB | 150+ tok/s |
| Qwen3.6 35B-A3B | Runs on GPU | Q8_0 | 38.2 GB | 150+ tok/s |
| gpt-oss-20b | Runs on GPU | Q8_0 | 23.2 GB | 128 tok/s |
| Ministral 3 3B | Runs on GPU | Q8_0 | 4.8 GB | 122 tok/s |
| Gemma 4 26B A4B | Runs on GPU | Q8_0 | 27.9 GB | 122 tok/s |
| Gemma 4 E4B | Runs on GPU | Q8_0 | 9.3 GB | 103 tok/s |
| Qwen3.5 9B | Runs on GPU | Q8_0 | 11.3 GB | 48 tok/s |
| Llama 4 Scout | Runs on GPU | Q4_K_M | 68.1 GB | 48 tok/s |
| Gemma 4 12B | Runs on GPU | Q8_0 | 13.8 GB | 39 tok/s |
| Phi-4 14B | Runs on GPU | Q8_0 | 17.7 GB | 31 tok/s |
| Mistral Small 3.1 24B | Runs on GPU | Q8_0 | 27.3 GB | 19 tok/s |
| Qwen3.6 27B | Runs on GPU | Q8_0 | 29.8 GB | 17 tok/s |
| Qwen3.8 27B | Runs on GPU | Q8_0 | 30.6 GB | 17 tok/s |
| Granite 4.2 30B | Runs on GPU | Q8_0 | 31.9 GB | 16 tok/s |
| Gemma 4 31B | Runs on GPU | Q8_0 | 33.8 GB | 15 tok/s |
| Hermes 4.3 36B | Runs on GPU | Q8_0 | 40.5 GB | 13 tok/s |
| Llama 3.3 70B | Runs on GPU | Q6_K | 61.0 GB | 8.5 tok/s |
| Qwen3.8-Flash-Next | Lower quality fit | IQ2_M | 60.5 GB | 150+ tok/s |
| gpt-oss-120b | Lower quality fit | Q3_K_M | 58.3 GB | 150+ tok/s |
| Qwen3.5 122B-A10B | Lower quality fit | Q3_K_M | 60.9 GB | 101 tok/s |
| GLM-5.3-Flash | Too big | Q4_K_M | 196.0 GB | — |
| DeepSeek V4.1-Flash | Too big | Q4_K_M | 335.8 GB | — |
OllamaLM StudioJanGPT4AllAnythingLLMllama.cppMLX LMKoboldCppText Generation WebUILocalAI