Back to browse
GitHub Repository

BitBop: latent-free ternary training of small language models

0 starsPython

Latent-free ternary LLM training

by Feyd·Jul 13, 2026·1 point·1 comment

AI Analysis

●●●BangerBig BrainWizardry

Fits a 325M model in 6GB VRAM where STE and float baselines crash.

Strengths
  • Eliminates full-precision latent weights, storing only flip momentum for ternary body.
  • Achieves parity with GPT-2-124M on BabyLM using strictly ternary weights.
  • Reference implementation runs on consumer RTX 3060, proving accessibility.
Weaknesses
  • No custom int2 kernel yet, so no actual inference speedup despite smaller storage.
  • Results limited to 125M-325M scale; asymptotic behavior at larger sizes unknown.
Category
Target Audience

ML researchers and engineers working on model quantization

Similar To

BitNet · Microscaling formats (MX)

Similar Projects