The all-in-one local AI studio for your desktop: chat, image and video generation and a coding agent in one free, open source app. Windows and Linux. No Docker, no terminal, no cloud required.
TypeScript agentai-chatai-studioall-in-onecoding-agentdesktop-appfluximage-generationlinuxllmlocal-ailocal-llmopen-sourceprivacyself-hostedstable-diffusionuncensoreduncensored-aivideo-generationwindows
/**
* The chat catalogue, read on a machine with one memory pool (a Mac).
*
* The catalogue in api/discover.ts is written from a PC: "fits a 12 GB card",
* "for high-VRAM setups", "Runs on your PC". An Apple Silicon Mac has no card
* and no VRAM of its own; processor and graphics unit share one pool, and a
* model takes its share of that pool. So the same numbers are said as memory
* here, and as FREE memory: a model written up for a 12 GB card needs about
* 12 GB that nothing else is using, which is not the same as a 12 GB Mac.
*
* Everything in this file is a rewrite at the point of display. With
* `shared` false every function returns what it was given, so a PC reads the
* catalogue exactly as before.
*
* NVFP4 and MXFP8: the catalogue itself now says what these Ollama tags are
* (MLX builds for Apple Silicon, see the note in api/discover.ts), so they
* need no rewrite here and read the same on every machine.
*/
import type { VramFit } from './vram-fit'
/**
* The share of a Mac's memory one model can take. macOS does not hand the
* whole pool to the graphics unit, and the system and the app live in the
* same pool. 0.7 is an estimate, not a measurement: on 16 GB it leaves about
* 11 GB, on 32 GB about 22 GB. It only moves the "runs on this Mac" hints;
* no download is ever held back by it.
*/
export const SHARED_MEMORY_MODEL_SHARE = 0.7
/** The memory a chat model is compared with: the card on a PC, a share of the pool on a Mac. */
export function chatModelBudgetGb(shared: boolean, graphicsGb: number | null, ramGb: number | null): number | null {
if (!shared) return graphicsGb
const pool = ramGb ?? graphicsGb
return pool ? Math.round(pool * SHARED_MEMORY_MODEL_SHARE) : null
}
// Order matters: the specific sentences go before the general words.
const REWRITES: ReadonlyArray<[RegExp, string]> = [
[/\bfits an? (\d+) ?GB card\b/g, 'fits in $1 GB of free memory'],
[/\bfits (\d+) ?GB (?:GPUs|VRAM)\b/gi, 'fits in $1 GB of free memory'],
[/\bruns on an? (\d+) ?GB card\b/g, 'runs in $1 GB of free memory'],
[/\bruns on (\d+) ?GB GPUs\b/g, 'runs in $1 GB of free memory'],
[/\bon an? (\d+) ?GB card\b/g, 'in $1 GB of free memory'],
[/\bon (\d+) ?GB cards\b/g, 'in $1 GB of free memory'],
[/\b([Ff])or (\d+) ?GB(\+?) cards\b/g, '$1or $2 GB$3 of free memory'],
[/\bfrom (\d+) ?GB VRAM\b/g, 'from $1 GB of free memory'],
[/\b([Ff])or high-VRAM setups\b/g, '$1or Macs with a lot of memory'],
[/\bHigh-VRAM setups\b/g, 'Macs with a lot of memory'],
[/\bhigh-VRAM setups\b/g, 'Macs with a lot of memory'],
[/\bbig VRAM only\b/g, 'a lot of memory only'],
[/\blow VRAM quant\b/g, 'low memory quant'],
[/\bVRAM\/RAM\b/g, 'memory'],
[/\bRAM\/VRAM\b/g, 'memory'],
[/\b(?:small|tiny) GPUs\b/g, 'Macs with little memory'],
[/\bfor edge GPUs\b/g, 'for small machines'],
[/\bVRAM\b/g, 'memory'],
]
/** A catalogue sentence as a Mac reads it. Unchanged where `shared` is false. */
export function sharedMemoryWording(text: string, shared: boolean): string {
if (!shared || !text) return text
return REWRITES.reduce((t, [from, to]) => t.replace(from, to), text)
}
/**
* What is left of PC hardware in a sentence after the rewrite. The guard test
* runs the whole catalogue through this, so a new entry that names a card in
* a way the table above does not know turns the test red instead of reaching
* a Mac. "model card" is a document, not hardware.
*/
export function pcHardwareLeft(text: string): string[] {
const hits = text.match(/\b(?:\d+ ?GB cards?|graphics cards?|VRAM|RTX[ -]?\d*|Blackwell|H100|MI300X|datacenter GPUs|your PC|your GPU)\b/g)
return hits ?? []
}
export interface FitWords { label: string; title: string }
const FIT_WORDS_SHARED: Record<VramFit, FitWords> = {
fits: { label: 'Runs on this Mac', title: "Fits in this Mac's memory with room to spare. Fast." },
tight: { label: 'Tight fit', title: 'Barely fits. With other apps open it can slow down.' },
big: { label: 'Needs more memory', title: 'Bigger than the memory this Mac can give a model. It may load very slowly or not at all.' },
unknown: { label: '', title: 'Memory not detected yet.' },
}
/** The verdict on a chat tile for a machine with one memory pool. */
export function sharedFitWords(fit: VramFit): FitWords {
return FIT_WORDS_SHARED[fit]
}