h34v7/Ternary-Qwen3.5-1B

Research - WIP

The idea is 0.8B > 1.6B bits of knowledge 1B > 1.58B bits of knowledge

distill 0.8B bf16 > 1B ternary

1.6B bits of knowledge to 1.58B frame

10x smaller memory, faster inference, same knowledge.

Will work? idk...