BitTern / CAT-Q: Cost-efficient and Accurate Ternary Quantization for LLMs
I stumbled across this cool paper, link below. They're using back propagation on the weights to progressively fit ternary using layer by layer reconstruction error.
It's surprising it works so well using a much smaller dataset than Unsloth, 512 x 2048, so 1 million tokens, and the loss being minimized is not KL divergence on the text generation, but reconstruction error inside the model.
The way it learns to scale and shift blocks of parameters - learnable modulation - seems conceptually similar to Unsloth, but rather than then applying hard quants, backpropping steadily more aggressive soft ternerizaition using layer reconstruction loss, and that seems to be the novel bit.
In the results table, something around the Llama2-70B size looks interesting, scoring well and not dropping much across the reasoning benchmarks, so it would become viable on a 24GB card. If that's the same for other big models that could be a useful unlock. But obviously the ancient Llama2-70B was eclipsed by 27B / 35B Qwen 3.6.
But more excitingly, maybe it makes something huge like MiniMax-M3 428B A23B viable on a 128GB RAM system, especially as prefill will get a lot faster in ternary?
Making your car engine lighter with an angle grinder gets better every day!
Paper:
https://arxiv.org/html/2606.26650v1
Not yet open source:
