Calculator
Which Mac runs which LLM, and how fast
Two numbers decide everything about running a model locally on Apple Silicon. Whether it fits in unified memory, which is arithmetic on the parameter count. And how fast it generates, which is memory bandwidth divided by the size of the weights. Pick your Mac and your model below and the calculator does both.
Yes. Llama 3.3 70B at Q4_K_M runs on M4 Pro with 64 GB.
Expect roughly 5.7 tokens/sec of generation. The weights are 42.8 GB, the 8K tokens context adds about 1 GB, and 53.2 GB of your 64 GB is spoken for once macOS is accounted for.
The two formulas
Memory. Weights in gigabytes are parameters in billions times bits per weight divided by eight. Q4_K_M averages about 4.85 bits once GGUF metadata is counted, Q8_0 about 8.5, FP16 exactly 16. Add 15% for runtime overhead, add the KV cache for your context length, then leave 3 GB for macOS. A model that fits with nothing spare will swap to SSD the moment you paste a long document, and swapping costs far more than any quantization ever saves you.
Speed. Generating one token means reading every active weight out of memory exactly once, so bandwidth divided by weight size sets the ceiling. Sitting on top of that is a fixed cost per token that does not shrink with the model: sampling, kernel launches, framework overhead. Seconds per token is the sum of the two. Fitted against the six models we measured on a base M4, that comes out at 91.5% of theoretical bandwidth plus 3.9 ms of overhead, and reproduces all six figures to within 4%.
That second term is what most calculators leave out, and it is why their numbers stop making sense at the small end. Bandwidth alone predicts 525 tokens/sec for Llama 3.2 3B on an M5 Ultra. Ours predicts 177, because once the weights are down around two gigabytes the fixed cost per token, not the memory bus, is what you are up against. Method and raw JSON for the underlying measurements are on the benchmarks page.
Where this math is wrong
Worth knowing before you spend money on the strength of a number from a web page.
- The overhead term is held constant across chips. In reality it almost certainly falls on faster silicon, which means the figures for Max and Ultra chips are more likely to be low than high. We would rather under promise a machine you are about to buy.
- Prompt processing is a different number entirely. Prefill is compute bound, not bandwidth bound, and it varies by roughly 2x between model families at the same size. We measured Qwen 2.5 7B at 1,130 prompt tokens/sec against 587 for Llama 3.1 8B on identical hardware. If you feed the model whole documents and want short answers, that gap matters more than the generation figure everyone quotes.
- Every row except the M4 is calculated, not measured. The formula is validated on one chip so far. It matches what the community reports on the rest of the lineup, but we label estimates as estimates.
- Binned chips are slower than their name suggests. An M4 Max with a 32 core GPU runs at 410 GB/s and one with 40 cores at 546. Same name on the box, 33% difference in generation speed.
Memory and speed for every model, at Q4_K_M
Minimum memory assumes an 8K context and 3 GB left for the system. Speeds are generation, in tokens per second. A dash means the model does not fit on that chip at its maximum memory configuration.
| Model | Weights | Min memory | M4 | M4 Pro | M4 Max (40-core GPU) | M5 Ultra |
|---|---|---|---|---|---|---|
| Llama 3.2 3B | 1.9 GB | 8 GB | 46.4 | 86 | 129 | 177 |
| Mistral 7B | 4.4 GB | 16 GB | 22.9 | 46.8 | 79 | 127 |
| Qwen 2.5 7B | 4.6 GB | 16 GB | 21.8 | 44.8 | 76 | 124 |
| Llama 3.1 8B | 4.9 GB | 16 GB | 20.8 | 42.9 | 74 | 120 |
| DeepSeek R1 Distill 8B | 4.9 GB | 16 GB | 20.8 | 42.9 | 74 | 120 |
| Qwen 2.5 14B | 9 GB | 16 GB | 11.7 | 25.1 | 45.8 | 83 |
| Qwen 2.5 32B | 19.9 GB | 32 GB | 5.4 | 12 | 22.9 | 45.5 |
| Llama 3.3 70B | 42.8 GB | 56 GB | — | 5.7 | 11.2 | 23.3 |
| Qwen 2.5 72B | 44.1 GB | 56 GB | — | 5.5 | 10.9 | 22.7 |
| Mistral Large 2 | 74.6 GB | 96 GB | — | — | 6.5 | 13.9 |
| Llama 3.1 405B | 245.5 GB | 288 GB | — | — | — | 4.4 |
| DeepSeek V3 | 406.8 GB | 480 GB | — | — | — | 41.1 |
Speeds for each chip assume its largest memory configuration: M4 at 32 GB, M4 Pro at 64 GB, M4 Max (40-core GPU) at 128 GB, M5 Ultra at 512 GB.
Apple Silicon memory bandwidth, M1 to M6
The single most useful table if you are choosing hardware, because bandwidth is what you are actually buying. Figures are Apple’s own published specifications, verified 2026-09-11.
| Chip | Bandwidth | Memory options | Found in |
|---|---|---|---|
| M1 | 68 GB/s | 8, 16 GB | MacBook Air, 13-inch MacBook Pro, Mac mini, iMac (2020-2021)16 GB ceiling makes 14B the hard limit, and only at Q4. |
| M1 Pro | 200 GB/s | 16, 32 GB | MacBook Pro 14/16-inch (2021) |
| M1 Max | 400 GB/s | 32, 64 GB | MacBook Pro 14/16-inch (2021), Mac Studio (2022)Still one of the best value-per-GB/s machines on the used market. |
| M1 Ultra | 800 GB/s | 64, 128 GB | Mac Studio (2022) |
| M2 | 100 GB/s | 8, 16, 24 GB | MacBook Air, 13-inch MacBook Pro, Mac mini (2022-2023) |
| M2 Pro | 200 GB/s | 16, 32 GB | MacBook Pro 14/16-inch, Mac mini (2023) |
| M2 Max | 400 GB/s | 32, 64, 96 GB | MacBook Pro 14/16-inch, Mac Studio (2023) |
| M2 Ultra | 800 GB/s | 64, 128, 192 GB | Mac Studio, Mac Pro (2023)192 GB at 800 GB/s still holds 123B-class models at Q8. |
| M3 | 100 GB/s | 8, 16, 24 GB | MacBook Air, 14-inch MacBook Pro, iMac (2023-2024) |
| M3 Pro | 150 GB/s | 18, 36 GB | MacBook Pro 14/16-inch (2023)Bandwidth dropped 25% versus M1 Pro and M2 Pro. The weakest Pro tier for inference. |
| M3 Max (30-core GPU) | 300 GB/s | 36, 96 GB | MacBook Pro 14/16-inch (2023)The binned M3 Max. 300 GB/s against 400 on the full chip. |
| M3 Max (40-core GPU) | 400 GB/s | 48, 64, 128 GB | MacBook Pro 14/16-inch (2023) |
| M3 Ultra | 819 GB/s | 96, 256, 512 GB | Mac Studio (2025)512 GB was the first configuration able to hold a 671B MoE in memory. |
| M4 | 120 GB/s | 16, 24, 32 GB | MacBook Air, 14-inch MacBook Pro, Mac mini, iMac (2024-2025)The one row on this page we have measured directly. See /benchmarks. |
| M4 Pro | 273 GB/s | 24, 48, 64 GB | MacBook Pro 14/16-inch, Mac mini (2024)64 GB is the cheapest sensible way into 70B at Q4. |
| M4 Max (32-core GPU) | 410 GB/s | 36, 96 GB | MacBook Pro 14/16-inch (2024), Mac Studio (2025)The binned M4 Max, 410 GB/s against 546 on the full chip. |
| M4 Max (40-core GPU) | 546 GB/s | 48, 64, 128 GB | MacBook Pro 14/16-inch (2024), Mac Studio (2025) |
| M5 | 153 GB/s | 16, 24, 32 GB | MacBook Air, 14-inch MacBook Pro (2025-2026) |
| M5 Pro | 307 GB/s | 24, 48, 64 GB | MacBook Pro 14/16-inch, Mac mini (2026) |
| M5 Max (base) | 460 GB/s | 36, 48, 64, 128 GB | MacBook Pro 14/16-inch, Mac Studio (2026)Configurable to 614 GB/s. Check which one you are actually buying. |
| M5 Max (614 GB/s) | 614 GB/s | 48, 64, 128 GB | MacBook Pro 14/16-inch, Mac Studio (2026) |
| M5 Ultra | 1200 GB/s | 96, 256, 512 GB | Mac Studio (2026)1.2 TB/s and 512 GB. The fastest single box for local inference Apple has shipped. |
| M6 | 153 GB/s | 16, 24, 32 GB | Mac mini (2026)Up to 170 GB/s on higher configurations; 153 is the entry figure used here. |
Common questions
How much RAM do I need to run a 70B model on a Mac?
About 56 GB at Q4_K_M with a short context: roughly 43 GB of weights, 15% runtime overhead, a gigabyte or two of KV cache, and 3 GB left for macOS. That makes 64 GB the smallest practical configuration, which in the current lineup means an M4 Pro, M5 Pro, or any Max or Ultra chip. At Q8 the same model needs around 128 GB.
Does a faster GPU make local LLMs faster on a Mac?
Barely, for generation. Producing one token means reading every active weight out of memory once, so generation speed is set by memory bandwidth, not GPU cores. A 24 GB M4 with more GPU cores runs an 8B model at the same speed as a 16 GB M4 because both sit on the same 120 GB/s bus. Prompt processing is the exception: that stage is compute bound and does scale with cores.
How accurate are these speed estimates?
They reproduce our own measurements to within 4%. On a base M4 the model predicts 46.4 tokens/sec for Llama 3.2 3B against 46.7 measured, 20.8 for Llama 3.1 8B against 21.2, and 11.7 for Qwen 2.5 14B against 11.7. Those six models on that one chip are everything we have measured directly, so every other row is extrapolation from the same two constants rather than a reading off a machine. Treat them as planning numbers, accurate enough to choose hardware by and not precise enough to quote.
Why does the calculator reserve 3 GB for the system?
Because a model that fits with nothing to spare is a model that swaps. macOS, a browser, and whatever agent process is calling the model all need memory, and once the machine starts paging to SSD the generation speed collapses regardless of what the bandwidth math says.
Do mixture-of-experts models follow the same math?
Memory yes, speed no. An MoE model has to hold every expert in unified memory, so the total parameter count decides whether it fits. Per token it reads only the active experts, so generation speed follows the active parameter count. That is why a 671B MoE with 37B active can generate faster than a dense 70B while needing eight times the memory.
When the answer is a Mac you do not own
If the model you want needs 64 GB or more and the machine has to be reachable around the clock, buying is not the only option. We rent dedicated Apple Silicon with Ollama already installed and an OpenAI compatible endpoint already exposed, from a base M4 Mac mini at $79/mo up to a 256 GB Mac Studio. Same hardware, same math as this page, none of the hosting.