People ask whether a Mac Studio is a good machine for running language models locally, and the answer usually arrives as an opinion. This guide replaces the opinion with arithmetic. Apple publishes the memory sizes and memory bandwidth of each Mac Studio configuration. Our GPU memory calculator gives the size of a model's weights and key-value cache. Put the two together and you can say which models fit in which configuration, and what the highest possible generation speed is. The numbers are planning figures, not benchmarks: we do not own a Mac Studio, and nothing below was measured on one.
What Apple lists
The Mac Studio Technical Specifications page on apple.com (read on 2026-11-13) lists two chip families, each with memory sizes and a memory bandwidth figure. The memory is unified: the CPU and GPU share one pool, so a model is not split between "system RAM" and "VRAM" as it would be on a PC with a discrete graphics card.
| Chip configuration | Unified memory options | Memory bandwidth |
|---|---|---|
| M5 Max, 18-core CPU, 32-core GPU | 36 GB | 460 GB/s |
| M5 Max, 18-core CPU, 40-core GPU | 48 GB, 64 GB or 128 GB | 614 GB/s |
| M5 Ultra, 30-core CPU, 64-core GPU | 96 GB | 1.2 TB/s |
| M5 Ultra, 36-core CPU, 80-core GPU | 512 GB (the 256 GB option is listed too; see note below) | 1.2 TB/s |
One detail matters for planning: the larger memory sizes are tied to the larger chips. You cannot pair 128 GB with the 460 GB/s chip, and 512 GB is listed only for the M5 Ultra with the 36-core CPU and 80-core GPU. Apple's configure-to-order list gives 256 GB without naming a chip, so we show it as an available Ultra option and cannot say from the page which Ultra configurations accept it; check Apple's configurator before you plan around 256 GB. Prices on the page are loaded dynamically and are not quoted here, so check Apple for current pricing and availability before you decide.
Step 1: does the model fit?
Memory capacity decides whether a model runs at all. The method is the same as in our memory estimation guide:
weights (GB) = parameters (billions) × bits per weight ÷ 8
KV cache (GB) = layers × KV values per layer × bytes per value × context tokens ÷ 10⁹
Take Llama 3.3 70B at 4 bits with a 32,768-token context and one user. The weights are 70 × 4 ÷ 8 = 35.0 GB. The cache stores 80 layers × 2,048 values × 2 bytes = 327,680 bytes per token, which for 32,768 tokens is 10.7 GB. The model needs about 45.7 GB before any runtime overhead (see the KV cache guide for where the 2,048 comes from).
Unlike a dedicated GPU, a Mac must also run macOS and your other applications from the same pool. The Apple page does not say how much of the unified memory one inference process may use, so we use a plain planning assumption: keep 25% free, meaning weights plus cache should be at most 75% of the installed memory. That 25% is our assumption, not an Apple figure; adjust it for your own workload.
| Model | Precision | Weights | KV cache | Total | Smallest option that fits |
|---|---|---|---|---|---|
| Qwen 2.5 7B | 4-bit | 3.5 GB | 1.9 GB | 5.4 GB | 36 GB |
| Qwen 2.5 7B | 8-bit | 7.0 GB | 1.9 GB | 8.9 GB | 36 GB |
| Llama 3.3 70B | 4-bit | 35.0 GB | 10.7 GB | 45.7 GB | 64 GB |
| Llama 3.3 70B | 8-bit | 70.0 GB | 10.7 GB | 80.7 GB | 128 GB |
| Llama 3.1 405B | 4-bit | 202.5 GB | 16.9 GB | 219.4 GB | 512 GB |
| Llama 3.1 405B | 8-bit | 405.0 GB | 16.9 GB | 421.9 GB | none listed |
| DeepSeek-V3 671B | 4-bit | 335.5 GB | 2.3 GB | 337.8 GB | 512 GB |
| DeepSeek-V3 671B | 8-bit | 671.0 GB | 2.3 GB | 673.3 GB | none listed |
Weights use nominal parameter counts (for example 7B for Qwen 2.5 7B, whose model card gives about 7.6B), so real files are slightly larger. Model architectures come from the same presets the calculator uses (layers and KV dimensions from each model's published configuration; DeepSeek-V3 uses multi-head latent attention, which is why its cache is so small). Two cautions apply to the table. First, nominal "4-bit" is optimistic. Common 4-bit file formats store scales alongside the weights and average closer to 4.5 to 5 bits per weight; at 4.9 bits, as in our speed guide, the 70B weights become 42.9 GB, the total 53.6 GB, and the 75% rule asks for 71.5 GB of memory, which moves you from the 64 GB option to 128 GB. Second, the cache grows linearly with context: at 128K tokens the 70B cache alone is about 43 GB.
Step 2: how fast can it generate?
Fitting is necessary but not sufficient. Generating one token at a time is limited by how fast the machine can read the model from memory, as explained in the speed guide:
tokens per second ≤ memory bandwidth ÷ bytes read per token
For a dense model, the bytes read per token are the weights plus the cache. For DeepSeek-V3, a mixture-of-experts model with 37B active of 671B total parameters (see our MoE guide), only the active share is read, while the whole model still has to fit in memory. Using the bandwidth of the smallest option that holds each model:
| Model | Precision | Bytes read per token | Chip tier | Ceiling |
|---|---|---|---|---|
| Qwen 2.5 7B | 4-bit | 5.4 GB | 460 GB/s | about 85 tokens/s |
| Qwen 2.5 7B | 8-bit | 8.9 GB | 460 GB/s | about 52 tokens/s |
| Llama 3.3 70B | 4-bit | 45.7 GB | 614 GB/s | about 13 tokens/s |
| Llama 3.3 70B | 8-bit | 80.7 GB | 614 GB/s | about 7.6 tokens/s |
| Llama 3.1 405B | 4-bit | 219.4 GB | 1.2 TB/s | about 5.5 tokens/s |
| DeepSeek-V3 (37B active) | 4-bit | 20.8 GB | 1.2 TB/s | about 58 tokens/s |
The pattern is the useful part. Doubling memory capacity does not double speed: a 70B model on 128 GB reads the same bytes per token as on 64 GB, so speed is set by the chip's bandwidth, and 8-bit weights halve the ceiling because they double the bytes read. The mixture-of-experts model is the surprise in the table: it needs the biggest machine to fit, yet its ceiling is far above the dense 405B's because it reads only a small slice of its weights for each token.
These ceilings are not predictions. Real engines reach a fraction of peak bandwidth, and the mixture-of-experts figure assumes the best case in which only the 37B active parameters are read. We have no Mac measurements, so we cannot tell you what fraction to expect; treat any published benchmark that falls well below these ceilings as normal and any that exceeds them as a sign the test used a smaller model, a shorter context or batching.
Where the Mac Studio is the right shape, and where it is not
- Large models for one user. The 70B 8-bit and 405B 4-bit rows need 80 to 220 GB. A single consumer GPU such as the RTX 4090, which NVIDIA lists with 24 GB of GDDR6X, cannot hold any of them, and splitting across several cards brings the PCIe and power issues covered in our home 70B guide. One Mac Studio holds the whole model in a single address space. That is the real advantage of unified memory: capacity per machine.
- Small models, where speed is the goal. If the model fits in 24 GB, capacity is no longer the constraint and the ranking is decided by bandwidth and by software support. In that regime the capacity advantage disappears, and a discrete GPU with high bandwidth deserves a look.
- Prompt processing. Everything above concerns token generation. Reading a long prompt (prefill) is limited by compute rather than bandwidth, so a long document can take far longer to start answering than the generation ceilings suggest. We have not measured this on Apple hardware, so we make no claim about the size of the effect.
- Many users at once. Each user adds its own cache, and the table's single-user ceiling is shared. Server-style batching is where data-center GPUs and engines such as those in our inference server comparison are designed to excel.
A short checklist before you buy
- Pick the model and precision you actually want to run and compute weights plus cache at your real context length with the calculator.
- Divide by 0.75 (or your own headroom) and round up to a memory option that Apple lists.
- Look up that option's bandwidth in the table above and divide by the bytes read per token to get your ceiling. If the ceiling is below the speed you need, a bigger memory option will not fix it; only a higher-bandwidth chip or a smaller or lower-precision model will.
- Decide whether prompt processing time matters for your use. If you feed long documents, look for measured prefill figures on the exact configuration before buying.
What we have not verified
We did not benchmark any Mac. The memory share available to a model, the real fraction of bandwidth that inference software achieves on Apple's GPU, and prefill speed are all unmeasured here, and we say so rather than guess. If we obtain access to a machine we will replace these planning figures with measured ones and mark the date.
Sources
- Apple Mac Studio Technical Specifications (chips, memory options, memory bandwidth; read 2026-11-13)
- NVIDIA GeForce RTX 4090 (24 GB GDDR6X; read 2026-11-13)
- The Llama 3 Herd of Models (Dubey et al., 2024), Table 3
- DeepSeek-V3 README (671B total, 37B activated)
- Model layer counts and KV dimensions: the calculator presets documented on our methodology page.