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
//! `keep_alive` duration syntax: `5m` / `120s` / `1h` / `inf` (server default `10m`).
//! `None` means `inf` (never auto-unload).
use chrono::{DateTime, Utc};
use std::time::Duration;
pub fn parse(s: &str) -> Result<Option<Duration>, String> {
let t = s.trim().to_ascii_lowercase();
if t.is_empty() {
return Err(format!("invalid duration: {s}"));
}
if t == "inf" {
return Ok(None);
}
let (num, unit) = t.split_at(t.len() - 1);
let n: u64 = num.trim().parse().map_err(|_| format!("invalid duration: {s}"))?;
if n == 0 {
return Err(format!("invalid duration: {s}"));
}
let secs = match unit {
"s" => n,
"m" => n * 60,
"h" => n * 3600,
_ => return Err(format!("invalid duration: {s}")),
};
Ok(Some(Duration::from_secs(secs)))
}
pub fn default() -> Option<Duration> {
parse("10m").unwrap()
}
/// `inf` → None; otherwise now + duration.
pub fn unload_at(dur: Option<Duration>, now: DateTime<Utc>) -> Option<DateTime<Utc>> {
let d = dur?;
let chrono_dur = chrono::Duration::from_std(d).ok()?;
Some(now + chrono_dur)
}