Hacker News AIT2·71 pts
- Per-row bit allocation via Fisher information plus trellis-coded post-training quantization preserves math reasoning at ~2.3 bits per weight
- The 1.23 GB Qwen3.5-4B solves 330 of 540 exact-answer math problems, edging out the same-size Bonsai 2 Ternary 4B at 328
- Full recipe, KL evaluations and ~$34 GPU cost are published under the MIT license