M3 generation
Local LLMs on the M3 Ultra
The M3 Ultra moves 819 GB/s and takes up to 512 GB of unified memory. That combination runs everything up to DeepSeek V3 at Q4, at roughly 23.5 tokens/sec. Bandwidth sets the speed, memory sets the ceiling, and the tables below give both for every model.
Worth knowing. 512 GB was the first configuration able to hold a 671B MoE in memory.
Memory bandwidth
819 GB/s
Unified memory
96, 256, 512 GB
8B model at Q4
~102 tok/s
Found in: Mac Studio (2025). Specifications verified 2026-09-11against Apple’s published figures.
Mac Studio with the M3 Ultra for local LLMs
The Mac Studio is the desktop that carries the M3 Ultra, and for local inference it is a different class of machine from a Mac mini. The mini stops at the M4 Pro: 64 GB and 273 GB/s. The Studio with the M3 Ultra goes to 512 GB and 819 GB/s, which is 3x the bandwidth and therefore about that much faster on any model that fits both. At 512 GB it holds DeepSeek V3 at Q4_K_M, which no mini can.
Two things to check before ordering a Studio for this: memory is fixed at purchase, so size it for the largest model you expect to run, not the one you run today, and the price gap to a used RTX 3090 or a DGX Spark is on the hardware comparison. The calculator gives the memory and tokens per second for any model on this chip.
What the M3 Ultra runs, at Q4_K_M
Memory figures include a 3 GB reserve for macOS and an 8K context. Speeds come from a model fitted to our own measured benchmarks; the method and its error bars are on the calculator page.
| Model | Needs | 96 GB | 256 GB | 512 GB |
|---|---|---|---|---|
| Llama 3.2 3B | 8 GB | 171 tok/s | 171 tok/s | 171 tok/s |
| Mistral 7B | 16 GB | 110 tok/s | 110 tok/s | 110 tok/s |
| Qwen 2.5 7B | 16 GB | 106 tok/s | 106 tok/s | 106 tok/s |
| Llama 3.1 8B | 16 GB | 102 tok/s | 102 tok/s | 102 tok/s |
| DeepSeek R1 Distill 8B | 16 GB | 102 tok/s | 102 tok/s | 102 tok/s |
| Qwen 2.5 14B | 16 GB | 65 tok/s | 65 tok/s | 65 tok/s |
| Qwen 2.5 32B | 32 GB | 33.3 tok/s | 33.3 tok/s | 33.3 tok/s |
| Llama 3.3 70B | 56 GB | 16.4 tok/s | 16.4 tok/s | 16.4 tok/s |
| Qwen 2.5 72B | 56 GB | 16 tok/s | 16 tok/s | 16 tok/s |
| Mistral Small 4 | 88 GB | 10 tok/s | 10 tok/s | 10 tok/s |
| Mistral Large 2 | 96 GB | 9.7 tok/s | 9.7 tok/s | 9.7 tok/s |
| Mistral Medium 3.5 | 96 GB | 9.3 tok/s | 9.3 tok/s | 9.3 tok/s |
| DeepSeek V4 Flash | 208 GB | — | 4.3 tok/s | 4.3 tok/s |
| Llama 3.1 405B | 288 GB | — | — | 3 tok/s |
| DeepSeek V3 | 480 GB | — | — | 23.5 tok/s |
M3 Ultra against the M2 Ultra
Bandwidth went from 800 GB/s to 819 GB/s, which is 1.02x, and that ratio carries almost directly into generation speed at the same model and quantization. The memory ceiling went from 192 GB to 512 GB, which is the part that decides whether a model runs at all rather than how fast. That extra headroom is the stronger reason to upgrade, because speed you can wait out and memory you cannot.
Common questions
What is the largest model an M3 Ultra can run?
At Q4_K_M with 512 GB of unified memory, DeepSeek V3 is the largest of the major open models that fits, needing about 480 GB once macOS is accounted for. Anything larger has to drop to a smaller quantization, run across clustered machines, or move to a chip with a higher memory ceiling.
How fast is the M3 Ultra for local inference?
It has 819 GB/s of memory bandwidth, and generation speed on Apple Silicon is set almost entirely by bandwidth divided by the size of the weights. In practice that works out to roughly 102 tokens/sec on an 8B model at Q4 and 33.3 on a 32B, assuming they fit. That is about 1.02x the M2 Ultra at the same model size.
Is a Mac Studio with the M3 Ultra good for local LLMs?
Yes, and it is the reason the Mac Studio exists in this market: it is the only Apple desktop that ships the M3 Ultra with up to 512 GB of unified memory and 819 GB/s of memory bandwidth. A Mac mini tops out at the M4 Pro with 64 GB and 273 GB/s, so the Studio is the step up when a model does not fit in 64 GB or when you want 3x the generation speed on the same model.
How much memory should I order with an M3 Ultra?
Memory decides which models you can run at all, and it cannot be upgraded later, so it is the one specification worth overbuying. M3 Ultra ships in 96, 256, 512 GB. The 96 GB configuration caps you at the small end; 512 GB is what opens up the larger models in the table on this page.
Rent an M3 Ultra instead of buying one
We run this exact chip as a dedicated machine, from $363/mo billed annually or $571/mo billed monthly, with Ollama installed and an OpenAI compatible endpoint already exposed. Same silicon and the same numbers as this page, without the hosting, the static IP, or being your own on call.