Ternary variant (1.71 bits/weight, ~5.9GB). Matches Qwen3.6-27B FP16 (95.30) on GSM8K — near-zero math regression at 9.4x weight compression.

Highest among comparable base models.

1-bit variant (1.125 bits/weight, ~3.9GB, runs on iPhone 17 Pro Max). Math held to 92.8 at 14.2x compression.