Mac Studio for local LLMs: which M5 configuration to buy
The M5 Max and M5 Ultra Mac Studio for local LLMs: memory tiers, what Apple measured, which models fit, and how to use it from your laptop.
· Updated · 9 min read
The new Mac Studio is Apple's clearest pitch yet for running a local LLM on your own desk. Announced on 25 August 2026 with M5 Max and a new M5 Ultra chip, it goes up to 512GB of unified memory, and Apple pitches it directly at people who run large language models locally. If you are weighing a Mac Studio for local LLM work, the question is not whether it is fast. It is which configuration fits the models you actually want to run, and what Apple's headline AI figures do and do not tell you.
This piece sets out what Apple has published, as of 9 November 2026: the chips, the memory tiers and the bandwidth behind each one; what Apple's "up to 10.7x" claim measured, and the generation speeds Apple's own researchers later reported; how much memory common model sizes need; and how to use a Mac Studio from the laptop you carry, rather than only from the chair in front of it.
What Apple announced
Apple's press release put pre-orders on 25 August, with machines arriving from 22 September. The 512GB memory option was the exception: Apple said it was "coming in late October". As of 9 November, Apple's technical specifications page lists 512GB as a configure-to-order option for the top M5 Ultra.
The two chips, from that specifications page:
| M5 Max | M5 Ultra | |
|---|---|---|
| CPU | 18-core | 30-core, or 36-core |
| GPU | 32-core, or 40-core | 64-core, or 80-core |
| Unified memory | 36GB; 48, 64 or 128GB with the 40-core GPU | 96GB; 256GB or 512GB as upgrades (512GB needs the 80-core GPU; see note below the table) |
| Memory bandwidth | 460GB/s (32-core GPU), 614GB/s (40-core GPU) | 1.2TB/s |
| Source | Apple specs | Apple specs |
Two details in that table matter more for language models than anything else on the page. The memory you can choose is tied to the chip variant: 48GB and above on M5 Max needs the 40-core GPU, and 512GB needs the full M5 Ultra. And bandwidth roughly doubles from the top M5 Max to M5 Ultra (specs), which is where most of the difference in generation speed will come from (more on that below).
Apple's US starting prices in the announcement were $2,499 for M5 Max and $5,499 for M5 Ultra. Prices in other countries differ, and memory upgrades add to them; check Apple's store for your region on the day you buy.
What "up to 10.7x faster" actually measured
Apple's Mac Studio page says M5 Max delivers up to 10.7x faster "time to first token performance in LM Studio" than the M1 Max Mac Studio, and the press release puts it at 3.9x faster than M4 Max. For M5 Ultra the figures are up to 9.8x against M1 Ultra and up to 4x against M3 Ultra.
The footnotes on the Mac Studio page give the test: Apple measured in July 2026 on preproduction machines, timing the first token of an 8K-token prompt on a 14-billion-parameter model at 4-bit quantisation, in LM Studio v0.4.19+2. The M5 Max under test had the 40-core GPU and 128GB; the M5 Ultra had the 80-core GPU and 512GB.
That is a real and useful number, but it is one half of the job. Apple's own machine learning team explains the split in its post on MLX and the M5's Neural Accelerators (19 November 2025): "generating the first token is compute-bound", which is where the new Neural Accelerators in each GPU core help, while "generating subsequent tokens is bounded by memory bandwidth, rather than by compute ability". In that post, the base M5 gave up to about 4x faster time to first token than the M4, but 19 to 27% faster generation, which Apple attributes to 28% more memory bandwidth.
So read the launch figures this way:
- Time to first token is how long you wait before the answer starts. It grows with the prompt, and a coding agent sends long prompts: your files, the conversation, the tool output. This is what Apple's 10.7x and 9.8x measure, and it is where M5 changes the most.
- Generation speed, the tokens per second once the answer is flowing, follows memory bandwidth. The launch figures say nothing about it. The specifications show 1.2TB/s on M5 Ultra against 614GB/s on the top M5 Max, and Apple's press release describes the Ultra's bandwidth as 50 percent higher than before.
What Apple measured on M5 Ultra
The more useful numbers came later. On 15 October 2026 Apple's machine learning team published measurements of M5 Ultra against M3 Ultra, both Mac Studios with 512GB, using mlx_lm.benchmark from MLX LM with an 8,192-token context and 128 generated tokens. Unlike the launch figures, these include generation speed:
| Model (as Apple names it), precision | M3 Ultra, tokens/s | M5 Ultra, tokens/s | M3 Ultra, first token | M5 Ultra, first token |
|---|---|---|---|---|
| Qwen 3.8 27B, BF16 | 12.46 | 19.69 | 18.2s | 4.1s |
| Mistral Medium 3.5 128B, FP8 | 4.95 | 7.70 | 95.9s | 24.6s |
| Source | Apple ML Research |
The pattern is the one Apple's earlier post predicted. The first token arrives roughly four times sooner, thanks to the Neural Accelerators; generation is about 1.6 times faster, which Apple attributes to bandwidth rising from 819GB/s on M3 Ultra to 1.2TB/s. Both models were run in their native precision (BF16 and FP8), not at 4-bit, so a 4-bit version of the same model would need less memory and generate faster. Apple published no comparable figures for M5 Max in the Mac Studio.
How much memory a local LLM needs on a Mac Studio
On Apple silicon the GPU uses the same unified memory as everything else, so the memory you configure is the ceiling on the model you can load. The model's weights take most of it; the context (the conversation and files the model holds) and macOS itself need room on top.
The llama.cpp quantisation documentation gives sizes for Llama 3.1 at the common Q4_K_M 4-bit quantisation, a useful yardstick for any model of the same parameter count:
| Model (Llama 3.1) | Full size | At Q4_K_M | Source |
|---|---|---|---|
| 8B | 32.1GB | 4.9GB | llama.cpp |
| 70B | 280.9GB | 43.1GB | llama.cpp |
| 405B | 1,625.1GB | 249.1GB | llama.cpp |
Mixture-of-experts models change the speed picture but not the memory one: every expert still has to be in memory. Apple's MLX post notes that a 24GB MacBook Pro holds a 30B mixture-of-experts model at 4-bit with the whole workload under 18GB, which gives a sense of how far a modest machine goes.
Set against the Mac Studio's memory tiers from Apple's specifications, and the llama.cpp sizes above, our reading (an opinion, not a measurement) is:
- 36GB (M5 Max): comfortable for 8B to roughly 30B models at 4-bit, with room for a long context (Apple ran a 30B mixture-of-experts model at 4-bit in under 18GB). A 70B model at Q4_K_M (43.1GB of weights, per llama.cpp) does not fit.
- 48GB (M5 Max): a 70B model at 4-bit (43.1GB of weights, per llama.cpp) only just fits, with little left for context or anything else. Buy it for mid-sized models, not for 70B.
- 64GB and 128GB (M5 Max): a 70B model at 4-bit (llama.cpp: 43.1GB) with room for a working context; 128GB also allows higher-precision quantisations or two models loaded at once.
- 96GB (M5 Ultra): a 70B model at 4-bit with room for context, at roughly twice the bandwidth of the top M5 Max (specs).
- 256GB (M5 Ultra): a 405B dense model at 4-bit is out of reach: Llama 3.1 405B at Q4_K_M is 249.1GB of weights alone (llama.cpp). It suits several large models side by side, or very large mixture-of-experts models whose 4-bit size sits well below the limit.
- 512GB (M5 Ultra): the configuration Apple pitches for "frontier-class" models; a 405B dense model at 4-bit fits with room for context. In Apple's own runs, Mistral Medium 3.5 128B at FP8 took up to 158GB with a 65,536-token context.
For the specific models that fit each memory size today, from 16GB to 512GB, see our guide to the best local LLM by memory.
Clustering Mac Studios
Apple's press release says several Mac Studio systems can be linked over Thunderbolt 5 with RDMA (remote direct memory access) to pool memory, and that a cluster of four delivers "up to 3x faster AI inference than a single system". The footnote on the Mac Studio page gives the conditions: four M5 Ultra machines with 512GB each, connected in a full mesh, running mlx-lm summarising a 32K-token prompt on a 72-billion-parameter dense model, compared with one machine. The Mac Studio page also names exo as one tool for clustering.
Apple's October measurements show what clustering is for: a model too big for one machine. Muse Spark Open, an 816-billion-parameter mixture-of-experts model run at 4-bit (NVFP4), does not fit on a single Mac Studio; on four M5 Ultras, split with tensor parallelism, it generated 36.73 tokens per second at an 8,192-token context. In the same tests, four machines took Qwen 3.8 27B from 19.69 to 50.89 tokens per second.
For most people one machine with the right memory is the better buy. Clustering is for teams who need a model larger than one machine holds, or more speed than one machine gives.
Which runtimes use the M5 chips
Apple's own launch material points at two runtimes: LM Studio, which it used for the time-to-first-token figures, and MLX, Apple's open-source framework, which it used for image generation and for the clustering test (both per the footnotes on the Mac Studio page). The MLX post says MLX needs macOS 26.2 or later to use the M5's Neural Accelerators.
Ollama and llama.cpp also run on Apple silicon. Whether a given runtime version uses the M5's Neural Accelerators is something to check in that runtime's own release notes before relying on Apple's figures for it.
Use your Mac Studio from anywhere
A Mac Studio is a desktop. It sits at home or in the office; you are often somewhere else, with a laptop that cannot hold a 70B model. This is the problem Corewire is built for.
Corewire runs on the Mac Studio alongside the runtimes you already use: it adopts LM Studio, Ollama and MLX LM among others, and shares their models with your other computers over your own private, encrypted network. Each model gets one name, such as mac-studio/qwen3-coder:30b, that a coding agent on your laptop can use. The Claude Code guide shows the one command that starts Claude Code on a model on another of your machines; the Codex and OpenCode guides do the same for those agents.
You can also run the agent itself on the Mac Studio. The sessions guide shows how to start Claude Code or OpenCode on the Mac Studio from your laptop at a café, detach, and pick the same session up later from the train, the way tmux works, across your own devices. The agent and its model both run on the Mac Studio; your laptop only drives it.
Questions people ask
Is the Mac Studio good for running local LLMs?
It can hold very large models in a single machine, because its GPU uses up to 512GB of unified memory on M5 Ultra (Apple specs). Generation speed follows memory bandwidth, which is highest on M5 Ultra; how fast a given model feels depends on its size and quantisation.
Should I buy M5 Max or M5 Ultra for local AI?
Choose by memory first, then bandwidth. If the models you use fit in 128GB, an M5 Max with the 40-core GPU will run them; M5 Ultra roughly doubles bandwidth to 1.2TB/s (Apple specs), so the same model should generate faster, and it is the only way past 128GB.
Can a Mac Studio run a 70B model?
Yes, from 64GB upwards with comfortable room: a 70B model at Q4_K_M needs 43.1GB for its weights (llama.cpp). The 36GB base M5 Max cannot load it, and 48GB leaves little room for context.
Is Apple's 10.7x figure about tokens per second?
No. It is time to first token on an 8K-token prompt with a 14B 4-bit model in LM Studio (Apple). Generation speed depends mainly on memory bandwidth; Apple's October MLX measurements put M5 Ultra at about 1.6 times the generation speed of M3 Ultra.