Mac mini vs Mac Studio for running models

The two classes interleave on price and diverge on what they hold. Compare the top of the mini line against the bottom of the Studio line:

The Studio is $400 cheaper and 50 percent faster on paper — and it is the one that fails to load Llama-3.3-70B at Q4. The extra cores buy bandwidth, not memory. If the question is which class runs a given model, the answer is decided by the usable pool first and the bandwidth second.

Decode speed on a Mac is bandwidth over bytes per token, so each chip class prints its own ceiling. The number below is computed the same way on every row; it only applies where the row above it says the model loads.

Four speed grades: the base M6 at 170 GB/s, the M5 Pro at 307, the M5 Max at 460 to 614, the Ultra at 1,200. The M5 Max 36GB is the slow variant — Apple pairs that memory with the 32-core GPU only, so the cheapest Studio is not the 614 one.

The base M6 32GB is a real local-AI machine: every model in the 12–32B class loads at Q4. The ceiling is the M5 Pro 64GB — the largest pool Apple puts in a mini, 58.9GB usable.

One row decides the upgrade: Qwen3-Coder-Next misses the 48GB tier and loads only on the 64. Llama-3.3-70B at Q4 needs the 64GB mini too — every smaller one refuses it. The 48GB tier buys no model the 32GB lacks; what it buys is Qwen3.5-35B-A3B with headroom, and the Pro chip’s 307GB/s under it.

Past 59GB of model, no mini exists. The first one over the line, gpt-oss-120b, misses Q4 and Q3 on every mini; the first machine that holds it is the M5 Max 128GB at $5,099.

The frontier class is harsher than the Ultra 96 makes it look. MiniMax-M2, Qwen3-235B-A22B and DeepSeek-V4-Flash all need 141–173GB at Q4, so the 96GB Ultra at $5,499 refuses them too. The first box that runs this class is the 256GB Ultra at $9,499 — the same verdict the 128GB guide reaches from the other side.

One honest overlap: at the matched 64GB tier — mini against Studio — the Studio costs $600 more, runs the identical model list, and is exactly twice as fast: 9.4 against 4.7 tokens per second. If you have already settled on 64GB, the Studio is the speed buy and the mini is the quiet-desk buy.

Pick your model and the fit finder ranks every machine that holds it across both classes. The Grid carries today’s price for every row above.

The speed formula, the reserve rules and the calibration sources are on the methodology page. Every number above is reproducible from it.