Work out which Mac actually runs the model you want, and whether the expensive one is buying you speed or just headroom
Model you want to run
Workload
Raising the limit is a live sysctl, not a reboot setting:
sudo sysctl iogpu.wired_limit_mb=<megabytes>
It resets on reboot. Leave macOS at least 8 GB or you will trade a memory win for a swap penalty.
The lineup
| Machine | Memory | Usable | Bandwidth | Fits | Tok/s ceiling | Price USD | $ per tok/s |
|---|
Price is editable. Apple publishes a price only for the configurations marked below, so everything else starts blank - paste in what Apple's configurator quotes you and the $ per tok/s column fills in.
Disclaimer
Provided as is, with no warranty of any kind, express or implied, and no guarantee that any specification, price or figure here is accurate, complete or current. The tokens-per-second column is a roofline ceiling computed from published memory bandwidth, not a benchmark and not a measurement. Real throughput under llama.cpp, Ollama or MLX lands below it, by an amount that depends on the runtime, the quantization kernel, attention implementation and thermal headroom. This is not purchasing, financial or professional advice. Verify configurations and prices with Apple before spending money. Use entirely at your own risk.
Hardware specifications verified against Apple Newsroom on 28 August 2026. Nothing you enter here is transmitted anywhere.
As-is, no warranty. These apps are free under their listed license and run entirely in your browser. Use at your own risk — don't blame me if your PC catches fire, your dog runs away, or the math turns out wrong. Verify anything that actually matters. None of this is professional financial, medical, legal, or engineering advice.
Two separate questions get answered, because they have different answers and people conflate them constantly.
Does it fit? Memory has to hold the whole model, plus the KV cache, plus the runtime. That is driven by total parameters - in a mixture-of-experts model every expert must be resident even though most of them sit idle on any given token.
How fast will it feel? Generating a token means reading the weights that token actually
uses. That is driven by active parameters, and it is bandwidth-bound, so the ceiling is
simply bandwidth ÷ active weight bytes.
For a dense model those two numbers are the same, and a bigger machine mostly buys you the ability to run something else later. For a mixture-of-experts model they diverge by an order of magnitude, and that divergence is the entire buy decision.
The M5 Ultra takes 512 GB. That sounds like it runs anything, and it does not.
At Q4, a dense 405B model needs about 220 GB including its KV cache. It fits comfortably. But 203 GB of that is weights, and every byte of them is read for every token, so 1.2 TB/s divided by 203 GB gives a ceiling of roughly 6 tokens per second - and a ceiling is the best case, not the expected case. You would have spent five and a half thousand dollars to watch text arrive one word at a time.
Now take DeepSeek V3, which is 671B parameters with about 37B active. It needs about 350 GB resident, so it only runs on the biggest config. But it reads only 18.5 GB per token, which puts the ceiling near 65 tokens per second. Bigger model, far bigger memory requirement, eleven times the speed.
Same machine, same memory, same bandwidth. The difference is architecture, and no other configurator asks you about it.
Announced 25 August 2026, available 22 September, with the 512 GB Ultra following in late October.
| Machine | Unified memory | Bandwidth | Published starting price |
|---|---|---|---|
| Mac mini M6 | 16 GB, up to 32 GB | up to 170 GB/s | $899 |
| Mac mini M5 Pro | up to 64 GB | 307 GB/s | $1,699 |
| Mac Studio M5 Max | 36 GB, up to 128 GB | 460 to 614 GB/s | $2,499 |
| Mac Studio M5 Ultra | 96 GB, up to 512 GB | 1.2 TB/s | $5,499 |
Two details in that table are easy to miss and both cost real performance:
The M5 Max’s bandwidth is tied to the GPU, not the chip name. It starts at 460 GB/s and only reaches 614 GB/s with the 40-core GPU. Choosing the cheaper GPU quietly costs you about a quarter of your generation speed, on a machine you bought for generation speed.
Apple does not publish intermediate memory tiers or upgrade pricing. Only the configurations above are stated, so those are the only ones modelled here, and the price column in the tool is left blank wherever the published price does not correspond to that exact configuration. Paste in whatever Apple’s configurator quotes you and the dollars per token per second fills itself in. Inventing those numbers would have been worse than leaving them empty.
macOS does not let the GPU wire all of unified memory. The cap lives in the
iogpu.wired_limit_mb sysctl, and the default leaves a generous slice to the operating
system - the tool models that as about 75%, matching the convention used across the rest of
this site.
You can raise it:
sudo sysctl iogpu.wired_limit_mb=<megabytes>It takes effect immediately and resets on reboot. It is genuinely useful when a model just barely misses fitting. It is also a good way to make your machine miserable: starve macOS and it starts swapping, which is dramatically slower than the memory you just clawed back. Leave at least 8 GB. The tool offers 75%, 85% and 92% so you can see what the extra headroom actually buys before you go changing kernel parameters.
Decision log 2026-04-19 for this site says benchmarks are never estimated from specifications. This tool honours that: the tokens-per-second column is a roofline ceiling derived from published memory bandwidth, and it is labelled as such everywhere it appears. It is not a benchmark, it is not a measurement, and real throughput under llama.cpp, Ollama or MLX lands below it by an amount that depends on the runtime, the quantization kernel, the attention implementation and how hot the machine has got.
Treat it as an upper bound and a comparison tool between configurations, which is exactly what a buy decision needs. If you want the fuller latency picture including prefill and time-to-first-token, use the Inference Latency Estimator. For memory fit on NVIDIA hardware, use the GPU VRAM Calculator. To find a model that fits what you already own, try the Local Model Browser.
This tool is provided as is, with no warranty of any kind, express or implied, including no warranty of merchantability, fitness for a particular purpose, accuracy, completeness or currency. There is no guarantee that any specification, price or figure here is correct or current.
The tokens-per-second figures are theoretical ceilings computed from published bandwidth, not benchmarks and not measurements. This is not purchasing, financial or professional advice. Verify configurations, specifications and prices with Apple before spending money. To the fullest extent permitted by law, no liability is accepted for any loss or damage arising from use of this tool or reliance on its output. Use entirely at your own risk.
For informational purposes only. Not financial, medical, or legal advice. You are solely responsible for how you use these tools.