Tiny-LM efficiency track (Glint leaderboard). Auto-refreshes every 20s · …
ON THE BOARD (maintainer-verified, PR #76 merged):
#16 GoLLeM-v5 32M — eff 75.51 (BLiMP 73.77 / ARC-E 44.44 / byte_ppl 2.5386) ·
#21 GoLLeM-v5 16M — eff 74.49 ·
org
Fabryka AI.
CHASING #1 — JugnuLM-110M (eff 80.20). Our 32M carries a ~5% size-bonus edge; target for #1: BLiMP ~80 / ARC ~50 / byte_ppl ~2.0.
NOW TRAINING — 2×64M A/B: Qwen3 arch (RoPE-θ100k + RMSNorm + SwiGLU + QK-Norm) +
value residuals (ARC lever),
Muon vs AdamW (clean optimizer verdict) on identical data. The #1 shot. ·
board:
Glint Tiny-ML
