SYLIN

TTF ternary training

by oguzeldereli · Technique · Text & code

Claims ternary training at ~1 byte/weight: 27B model trained on one 24GB RTX 3090; non-commercial license; models can loop and lose factual recall

Evidence

  1. Nov 11
    Converted Qwen 3.8 27B GGUF (~6.7-8GB) matches full precision on GSM8K (0.895) but drops on HumanEval (0.84 vs 0.93)

Cost to run

Alternatives

Works with