Run 35B MoE models in ~2.5 GB of RAM. Weights stream from SSD; a trained router predicts loads a step ahead. Mac, iPhone, Android, Windows supported
Python
type Phase = "download" | "load";
interface Token {
tier: string;
taskId: string;
phase: Phase;
}
type Listener = () => void;
class AutoLoadStore {
private token: Token | null = null;
private listeners = new Set<Listener>();
get current(): Token | null {
return this.token;
}
subscribe(fn: Listener): () => void {
this.listeners.add(fn);
return () => {
this.listeners.delete(fn);
};
}
private emit(): void {
for (const l of this.listeners) l();
}
begin(tier: string, taskId: string): void {
this.token = { tier, taskId, phase: "download" };
this.emit();
}
claimCompleted(taskId: string): string | null {
if (!this.token || this.token.phase !== "download" || this.token.taskId !== taskId) return null;
const tier = this.token.tier;
this.token = { ...this.token, phase: "load" };
this.emit();
return tier;
}
consume(tier: string): boolean {
if (!this.token || this.token.phase !== "load" || this.token.tier !== tier) return false;
this.token = null;
this.emit();
return true;
}
cancel(): void {
if (this.token) {
this.token = null;
this.emit();
}
}
}
export const autoLoad = new AutoLoadStore();