klein4b-qat-ckpt — training checkpoint, not a usable model

Rolling checkpoint of an ongoing quantization-aware-training (QAT) experiment: the transformer of FLUX.2-klein-4B with its 100 block Linear layers trained as ternary (1.58-bit, absmean, group 128 on K) by distillation against the BF16 original.

ckpt/ultimo.pt is a raw PyTorch pickle (bf16 master weights in diffusers naming + torchao AdamW8bit state). It is overwritten every ~25 minutes, is not ternary-packed, and is not meant to be loaded for inference. Derived from FLUX.2-klein-4B, licensed Apache-2.0.