Add the MTP heads and the vision tower

Set quantization_config.bits to 2 in config.json

Drop fractional quantization_config.bits from config.json

Layer 47: interm_div instead of 6 bpw + bad-row override (2.27 bpw)

config: integer bits for Hub validation, true average kept as bits_per_weight

Model card, corrected config.json, repaired DFlash drafter config, eval artifacts