Mac StudioAvailable
Mac Studio M5 Ultra
30-core CPU / 64-core GPU, 96GB
$5,499at Apple Store
1,200 GB/s
Check priceMemory guide · 96GB unified memory
What can you run on 96GB unified memory?
12 of the 16 open models we track fit in 96GB unified memory without spilling into system RAM, the biggest being GPT-OSS 120B. Speeds range from about 23 to 120 tokens/sec depending on the model and machine. Prices & stock checked 2026-10-23. Prices change daily and some retailers block our checker, so always confirm on the store page.
Models that fit in 96GB unified memory
| Model | Needs at 4-bit | Mac Studio M5 Ultra1,200 GB/s |
|---|---|---|
| Qwen3.5 9Bdense · 9B active · chat, small agents | ~7 GB | Fits~89 tok/sQ8_0 |
| Gemma 4 12Bdense · 12B active · chat, vision | ~9 GB | Fits~69 tok/sQ8_0 |
| GPT-OSS 20BMoE · 3.6B active · reasoning, tools | ~13 GB | Fits~97 tok/sQ8_0 |
| Gemma 4 26B-A4BMoE · 3.8B active · fast general chat | ~18 GB | Fits~95 tok/sQ8_0 |
| Qwen3.8 27Bdense · 27B active · coding, vision, general | ~19 GB | Fits~33 tok/sQ8_0 |
| GLM-4.7-Flash 30B-A3BMoE · 3B active · coding agents | ~21 GB | Fits~103 tok/sQ8_0 |
| Gemma 4 31Bdense · 31B active · general, vision | ~21 GB | Fits~29 tok/sQ8_0 |
| Qwen3.6 35B-A3BMoE · 3B active · fast coding & chat | ~24 GB | Fits~103 tok/sQ8_0 |
| Llama 3.3 70Bdense · 70B active · general | ~46 GB | Fits~23 tok/sQ4_K_M |
| Qwen3-Next 80B-A3BMoE · 3B active · long context, fast | ~52 GB | Fits~120 tok/sQ4_K_M |
| Llama 4 Scout 109B-A17BMoE · 17B active · multimodal, long context | ~71 GB | Tight fit~60 tok/sQ4_K_M |
| GPT-OSS 120BMoE · 5.1B active · reasoning, tools | ~67 GB | Tight fit~109 tok/sMXFP4 (native) |
Each column shows the quant we'd use (8-bit when it fits comfortably and still runs 10+ tok/s, otherwise 4-bit) and the estimated decode speed for one user at ~8K context. How we estimate
Just out of reach: Nemotron 3 Super 120B-A12B (~78 GB at 4-bit), Qwen3.5 122B-A10B (~79 GB at 4-bit), Qwen3 235B-A22B (~151 GB at 4-bit).
1 tracked rig with 96GB unified memory
Also 96GB, but a discrete GPU: PNY NVIDIA RTX PRO 6000 Blackwell. It has a different usable share and bandwidth, so check its own page.
FAQ
What's the biggest model I can run on 96GB unified memory?
Of the models we track, GPT-OSS 120B is the largest that fits fully in 96GB unified memory (MXFP4 (native), about 67 GB including runtime and ~8K context).
What's the cheapest rig with 96GB unified memory?
Of the machines we track, the Mac Studio M5 Ultra at $5,499 (Apple Store, checked 2026-10-23).
Can I run Llama 3.3 70B on 96GB unified memory?
Yes, at 4-bit it needs about 46 GB including runtime and ~8K context.
How are fit and speed estimated?
Memory needed = weights + 1.5 GB + 5% of weights. Decode speed ≈ 1 ÷ (active weights in GB ÷ (efficiency × memory bandwidth) + a small fixed per-token overhead), with efficiency and overhead fitted per platform to 43 public llama.cpp/MLX benchmarks (median error about 5%). They're estimates, not benchmarks.