Observed arrival · 2026-09-11
Sane Labs, Teaching Small Models to Read
An independent language-model project training small models from scratch, with knowledge intended to come from supplied context rather than memorized weights.
○Open
⊠Login
$Paid
†Ads
✦Pretty
●Pro
◎Niche
◉Human
⚑Risk
ƒJS
Field notes
The project exposes unusually detailed training records, including model shape, token counts, hardware, precision, and weight size. Sane-47M and Sane-118M use an in-house BPE tokenizer and 1,024-token context, while the in-training Synth-2 expands to a listed 8k–32k context and uses 12 experts plus a shared expert. The page also separates released work from private or unfinished experiments rather than presenting every project as available.
Observed signals
Read the marks
Editorial observations of this landing page, not a rating.
○OpenPublic substance visible
✦PrettyNotable craft visible
●ProPolished or operationally mature
◎NicheUnusually specific use
One card from the complete issue