17 October 2026 · Jakub Mokříš
What a Mac Studio M5 Ultra with 256 GB of memory is good for
Large language models without the cloud, app builds, transcribing hours of recordings, or crunching data that won’t fit anywhere else. Here is what a 256 GB Mac Studio is genuinely useful for, and where it isn’t the right tool.
A Mac Studio with the M5 Ultra and 256 GB of unified memory is not something most people buy for a single project. For a few days, though, a surprisingly wide range of people can put it to good use. Here is what people come to us with and where the machine makes sense.
Large language models that won’t fit elsewhere
This is the main reason the machine is in our fleet. The GPU shares memory with the CPU, so a model doesn’t have to be split across several graphics cards. On our machine, gpt-oss-120b uses about 61 GB with a 32k context and generates around 120 tokens per second, which leaves plenty of room for bigger models.
In practice, you can try a model with a hundred billion parameters before paying for similar hardware, or run a model over your own documents without the data leaving a machine that only you are using. The numbers are in our own measurement and in our look at fifteen models.
Several models at once
256 GB isn’t only for one giant model. It often makes more sense to keep a fast coding model, a smaller model for triage and an embedding model for document search in memory at the same time. On a 32 or 64 GB machine they would have to take turns. Here they run side by side, and agent tools can switch between them without waiting for a model to load.
Fine-tuning on your own data
Apple’s MLX library supports LoRA fine-tuning, so you can adapt small and mid-sized models to your own style, terminology or output format. The large memory helps most with longer training texts and bigger models. Full training of large models from scratch is still a job for servers with NVIDIA GPUs.
Images and speech-to-text
Image generators such as Flux or Stable Diffusion run on Apple Silicon in apps like Draw Things or directly through MLX. For transcription there is Whisper, as whisper.cpp or mlx-whisper. This works well for an archive of interviews or podcasts: upload dozens of hours of audio, start a batch overnight and download the transcripts in the morning.
iOS and macOS app builds
Apps for Apple platforms can only be built on a Mac. The Mac Studio handles a large Xcode project, several simulators at once, and Unity or Unreal game builds for iPhone and Mac. You can also run your own GitHub Actions or GitLab CI runner on it for a while, which helps when you need faster builds before a release or want to check your app on the latest macOS. If you only need builds and not the memory, a Mac mini is often enough.
Video, 3D and audio
Final Cut Pro, DaVinci Resolve, Logic Pro and Blender with Metal rendering all run well on the Mac Studio. Remotely there is one catch. Exports, renders and batch conversions work great, but interactive editing over remote desktop won’t be smooth. Also keep in mind that footage has to be uploaded first, so your upload speed matters.
Data that won’t fit anywhere else
With 256 GB of memory, pandas, Polars or DuckDB can handle tables that a typical laptop would have to process in chunks. For a one-off analysis of a large export or a model recalculation over a weekend, this is often simpler than building cloud infrastructure.
What it isn’t good for
- CUDA-only software. Projects built solely for NVIDIA GPUs won’t run on a Mac, or only with changes.
- Windows programs. It’s a Mac running macOS, not a Windows server.
- Gaming over remote desktop and other latency-sensitive work.
- Remote desktop from Windows. It currently works only from a Mac. From Windows you connect over SSH; see our guides.
Try it
Daily and weekly rentals are currently free; only 30-day rentals are paid. Every rental also gets a free hour for setting up the connection. Just fill in the inquiry and tell us what you need the Mac for.