Hedroom for local AI
Running a language model on your own Mac depends mostly on whether it fits in memory. Hedroom estimates that before you download several gigabytes, compares it with what your Mac can really give, and helps make room if it is short.
Check before you load
In Tools › Local AI headroom you pick a model family and size, 4, 8 or 16 bits per parameter, and a context length from 4K to 128K. Hedroom adds up the weights, the context cache and the runtime, then compares the total with two numbers: what the GPU is allowed to use, and what is free right now. It answers one of four things, with the line that decided it:
- Fits.
- Tight.
- Make room.
- Won't fit.
The check and the reading of the GPU limit are free. Local AI headroom has the full list. Will a local AI model fit in your Mac's memory? explains the arithmetic and how to check it yourself without Hedroom.
Unified memory, and the GPU limit
On Apple Silicon the GPU shares memory with everything else, but macOS caps how much of it the GPU can hold. Apple does not publish the rule. Hedroom shows its own reading: about two-thirds of memory up to 36 GB, and three-quarters above. It says that this is its reading. A 16 GB Mac therefore has less for a model than its label suggests.
Make room
When your apps hold the memory a model needs, Make room suggests relaunching or quitting the apps that can give some back. It uses the same actions as the rest of Hedroom: asks first, goes through the safety checks, and writes a receipt. It does not clear caches, because that would not make room for anything.
Hedroom also notices Ollama, LM Studio, Jan and GPT4All when they are running, and shows what they use.
Raising the limit, only if you choose to
On a Mac with more than 24 GB, Hedroom can raise the GPU memory limit until you restart. It shows a warning first: a higher limit leaves less for macOS, and your Mac can slow down or become unstable if memory runs short. You confirm, and Hedroom never offers a limit that leaves macOS under 8 GB. When Hedroom quits it puts the limit back to what it was, unless you changed it yourself.
Raising the limit needs the same small helper as fan control, which you approve once in System Settings. Reading the limit does not. Raising it is part of Pro, and resetting it is never locked.
Honest limits
- Hedroom does not download, load or run a model, and it does not talk to a model app. It estimates memory only.
- No speed promises. A model that fits can still run slowly, and Hedroom does not estimate tokens per second.
- The estimate can be off by about 15%. The layer and head numbers are typical for a model family, not read from your file.
- On Macs with 24 GB or less, Hedroom does not offer a higher limit, because the most it would allow is no more than macOS already gives the GPU.
- The default GPU limit is an assumption that has not yet been checked on every memory size.
- Fan control and a higher limit have not yet been checked on a signed release build.
Try it, then decide
Monitoring and explanations are free, with no card and no account. Pro will have a 7-day free trial that needs a card, and you can cancel before it ends. Pro is a subscription, from $31.95 a year, or $7.95 a month, planned prices.