Marge pour l’IA locale
Gratuit à consulter, Pro pour modifier
Vérifiez si un modèle local tiendra dans la mémoire de votre Mac avant de le télécharger ou de le charger.
The problem
Downloading a local model is a long wait, and the failure comes at the end: it will not load, or it loads and your Mac starts swapping. The size on the download page is only part of the answer, because the context length you choose adds memory too.
Local AI headroom answers the question first. You pick the model size and context. Hedroom estimates what it needs, compares that with what your Mac can give a model right now, and names the line that decided the answer.
How it works
- Pick. Choose a model family and size, 4-, 8- or 16-bit, and a context from 4K to 128K. Hedroom names families and sizes only; it recommends no product.
- Estimate. It adds up the weights, the context cache (which grows with the context) and runtime overhead, each shown as "about". A 15% margin covers the fact that layer and head counts are typical for a family, not read from your file.
- Compare. The total is checked against the GPU memory limit and against what is free right now (free memory plus cached files; compressed memory does not count).
- Answer. Fits, Tight, Make room or Won't fit, with the line that decided it.
- Act, if you choose. Make room lists apps to relaunch or quit through the normal safety checks. On some Macs, Hedroom can raise the GPU limit until you restart, after a confirmation.
What you will see
A big verdict, "Fits.", "Tight.", "Fits, after making room." or "Won't fit.", with a sentence naming the line that decided it and the parts of the estimate: "Weights about … · context about … · runtime about … The estimate can be off by about 15%." A gauge shows what the model needs against two markers: what the GPU can use and what is free right now.
Beside it sits the GPU memory limit control, with a warning always on screen (less memory for macOS, your Mac can slow down or become unstable, it lasts until you restart), and a list of model apps Hedroom finds running: Ollama, LM Studio, Jan and GPT4All, with what they use. On a 16 GB Mac the limit control is off and says why.
A worked example
Example, not real data
A mid-size model on a 16 GB Mac
Say you want to try a Llama-family model at 8B parameters, 4-bit, with an 8K context. Hedroom estimates weights of about 4.8 GB, context of about 1.1 GB and runtime of about 0.8 GB: about 6.7 GB, or about 7.7 GB with the margin.
macOS lets the GPU use about two-thirds of a 16 GB Mac's memory by default, around 10.7 GB, so the GPU limit is not the problem. But say Chrome and Slack are open and free memory plus cached files reads about 7 GB. The low estimate fits and the high one does not, so the verdict is Tight, and the Make room list shows which apps hold the most. Hedroom promises no speed: it estimates memory only.
Details and limits
Ce qu’il fait
- Vous choisissez une famille et une taille de modèle, 4, 8 ou 16 bits, et une longueur de contexte de 4K à 128K.
- Estime ce dont il a besoin (poids, cache de contexte et exécution), puis compare avec ce que le GPU peut utiliser et ce qui est libre en ce moment.
- Répond « Ça tient », « Juste », « Faire de la place » ou « Ne tient pas », avec la ligne qui a tranché.
- « Faire de la place » suggère de relancer ou de quitter les apps qui occupent de la mémoire, via les actions normales de Hedroom.
- Repère Ollama, LM Studio, Jan et GPT4All quand ils tournent, et dit ce qu’ils utilisent.
- Peut relever la limite de mémoire GPU jusqu’au redémarrage, après une confirmation qui précise ce qu’il reste pour macOS.
What it will not do
- Télécharger, charger ou exécuter un modèle, ou parler à une app de modèle.
- Promettre de la vitesse. Il n’estime que la mémoire.
- Proposer une limite qui laisse moins de 8 Go à macOS.
- Relever la limite sans votre confirmation. Quand Hedroom quitte, il remet la limite à sa valeur précédente, sauf si vous l’avez modifiée vous-même.
Limits
- L’estimation peut être inexacte d’environ 15 %. Les nombres de couches et de têtes sont typiques d’une famille de modèles, pas lus dans votre fichier.
- macOS ne publie pas sa limite GPU par défaut. Hedroom affiche sa propre interprétation : environ les deux tiers de la mémoire jusqu’à 36 Go, les trois quarts au-delà.
- Sur les Mac de 8, 16, 18 ou 24 Go, Hedroom ne propose pas de limite plus élevée, car le maximum qu’il autoriserait n’est pas supérieur à ce que macOS accorde déjà au GPU.
- Relever la limite nécessite le même outil d’assistance que le contrôle des ventilateurs, qui demande une approbation dans Réglages Système. Une limite plus élevée laisse moins de mémoire à macOS, et votre Mac peut ralentir ou devenir instable si la mémoire vient à manquer. Cela dure jusqu’au redémarrage.
Gratuit ou Pro
La vérification et la lecture de la limite de mémoire GPU sont gratuites. Relever la limite fait partie de Pro. La réinitialisation n’est jamais verrouillée.
Raising the limit changes how your Mac behaves, which is why it is Pro.
Questions people ask
Does Hedroom download or run the model?
No. It never downloads, loads or runs a model, and never talks to a model app. It estimates memory and notices which model apps are running.
How accurate is the estimate?
It can be off by about 15%. The weights come from a simple formula, and layer and head counts are typical for the family, not read from your file. The verdict checks both the estimate and the estimate plus margin, so a borderline case reads Tight rather than Fits.
Is raising the GPU memory limit safe?
It leaves less memory for macOS, and your Mac can slow down or become unstable if memory runs short. Hedroom never offers a limit that leaves macOS less than 8 GB, shows how little is left before you confirm, puts the limit back when it quits, and the setting goes away at restart anyway.
Does it need the helper?
Checking is free and needs no helper. Raising or resetting the limit uses the same helper as fan control, which requires approval in System Settings. Approving the helper has the steps.