What fits in 96 GB of unified memory
Data updated . Last reviewed the same day, when this catalog snapshot was written.
2 catalog machines have 96 GB unified. Models are ranked against the Mac Studio M3 Ultra 96GB, the tightest usable pool at this sticker size (67.2 GB). The tok/s estimate uses Mac Studio M5 Ultra 96GB bandwidth (1200 GB/s) and is not a benchmark. 4 of these fits do not clear the 64 GB tier.
Models
| Model | Best quant | Estimate | Headroom | Est. tok/s |
|---|---|---|---|---|
| Llama 3.2 1B Instruct | FP16 | 3.0 GB | 64.2 GB | 485 tok/s |
| Llama 3.2 3B Instruct | FP16 | 7.7 GB | 59.5 GB | 187 tok/s |
| Phi-3.5 mini | FP16 | 9.2 GB | 58.0 GB | 157 tok/s |
| Phi-4 mini | FP16 | 9.2 GB | 58.0 GB | 156 tok/s |
| Gemma 3 4B | FP16 | 10.3 GB | 56.9 GB | 140 tok/s |
| Mistral 7B Instruct v0.3 | FP16 | 17.4 GB | 49.8 GB | 82.8 tok/s |
| Qwen2.5 7B Instruct | FP16 | 18.3 GB | 48.9 GB | 78.8 tok/s |
| Gemma 4 E4B | FP16 | 19.2 GB | 48.0 GB | 75.0 tok/s |
| DeepSeek R1 Distill Llama 8B | FP16 | 19.3 GB | 47.9 GB | 74.7 tok/s |
| Llama 3.1 8B Instruct | FP16 | 19.3 GB | 47.9 GB | 74.7 tok/s |
| Qwen3 8B | FP16 | 19.7 GB | 47.5 GB | 73.3 tok/s |
| Ministral 3 8B | FP16 | 21.4 GB | 45.8 GB | 67.3 tok/s |
| Gemma 2 9B | FP16 | 22.2 GB | 45.0 GB | 64.9 tok/s |
| Gemma 4 12B | FP16 | 28.7 GB | 38.5 GB | 50.2 tok/s |
| Gemma 3 12B | FP16 | 29.2 GB | 38.0 GB | 49.2 tok/s |
| Mistral Nemo 12B | FP16 | 29.4 GB | 37.8 GB | 49.0 tok/s |
| Ministral 3 14B | FP16 | 33.5 GB | 33.7 GB | 43.0 tok/s |
| Phi-4 | FP16 | 35.2 GB | 32.0 GB | 40.9 tok/s |
| Qwen3 14B | FP16 | 35.4 GB | 31.8 GB | 40.6 tok/s |
| Qwen2.5 14B Instruct | FP16 | 35.4 GB | 31.8 GB | 40.6 tok/s |
| gpt-oss-20b | FP16 | 50.2 GB | 17.0 GB | 167 tok/s |
| Mistral Small 3.2 24B | FP16 | 57.6 GB | 9.6 GB | 25.0 tok/s |
| Gemma 4 26B-A4B | FP16 | 61.9 GB | 5.3 GB | 158 tok/s |
| Gemma 2 27B | FP16 | 65.3 GB | 1.9 GB | 22.0 tok/s |
| Gemma 3 27B | FP16 | 65.8 GB | 1.4 GB | 21.9 tok/s |
| Qwen3.6 27B | FP16 | 66.7 GB | 0.5 GB | 21.6 tok/s |
| Qwen3.8 27B | FP16 | 66.7 GB | 0.5 GB | 21.6 tok/s |
| DeepSeek V4 Flash | Q8 | 4.5 GB | 62.7 GB | 7053 tok/s |
| GLM-4.7-Flash | Q8 | 32.3 GB | 34.9 GB | 392 tok/s |
| Qwen3 30B-A3B | Q8 | 33.3 GB | 33.9 GB | 376 tok/s |
| Gemma 4 31B | Q8 | 34.0 GB | 33.2 GB | 36.8 tok/s |
| Qwen3 32B | Q8 | 37.0 GB | 30.2 GB | 34.5 tok/s |
| DeepSeek R1 Distill Qwen 32B | Q8 | 37.0 GB | 30.2 GB | 34.5 tok/s |
| Qwen2.5 32B Instruct | Q8 | 37.0 GB | 30.2 GB | 34.5 tok/s |
| Qwen3.6 35B-A3B | Q8 | 37.1 GB | 30.1 GB | 390 tok/s |
| Mixtral 8x7B | Q8 | 50.7 GB | 16.5 GB | 87.5 tok/s |
| gpt-oss-120b | Q8 | 63.7 GB | 3.5 GB | 434 tok/s |
| Llama 3.3 70B Instruct | Q4 | 42.5 GB | 24.7 GB | 28.2 tok/s |
| DeepSeek R1 Distill Llama 70B | Q4 | 45.2 GB | 22.0 GB | 28.2 tok/s |
| Qwen2.5 72B Instruct | Q4 | 46.7 GB | 20.5 GB | 27.3 tok/s |
| Llama 4 Scout | Q4 | 65.4 GB | 1.8 GB | 117 tok/s |
Machines at 96 GB
- Mac Studio M3 Ultra 96GB
67.2 GB usable · $3,999
- Mac Studio M5 Ultra 96GB
67.2 GB usable · $5,499
Questions
- How many models fit in 96 GB of unified memory?
- 41 of 49 fit on the tightest config at this size, the Mac Studio M3 Ultra 96GB, which keeps 67.2 GB usable.
- Which catalog machines have 96 GB unified?
- Mac Studio M3 Ultra 96GB (70% usable), Mac Studio M5 Ultra 96GB (70% usable).
- Why is the ranking based on the tightest config?
- So a model listed as fitting works on every 96 GB machine here, including ones that reserve more memory for the OS. A looser machine has more headroom than the table shows.