2.2.0

Model bump: the shipped weights are now Prism ML's ternary build of Qwen3.8-27B. A re-download is coming (~8 GB, against ~13); the old snapshot can be freed with hf cache rm. The new default is nathansutton/Qwen3.8-27B-Ternary-Bonsai-2-DFlash2-MLX: the same model with every projection Hadamard-rotated and stored at 2 bits (levels {−s, 0, +s}), repacked text-only with the base tokenizer and the DFlash2 drafter bundled. 7.2 GB resident instead of 12.3, and on a dense model those 5 GB are context: the governor's window on a 24 GB Mac goes from ~56k tokens to ~150k. Decode holds: 64 tok/s drafted and 21 serial on the M4 Pro (the 3-bit: ~60 / 18). The cost is quality at the margin: teacher-forced perplexity on code is 4.49 against the 3-bit's 3.99 (+12%), while all six private eval tiers tie (56/56, no task flipped either way). The 3-bit quant stays one flag away: --model nathansutton/Qwen3.8-27B-UD-Q3_K_XL-DFlash2-MLX.

What made it run at that speed rather than as a port: the rotated weights need their transform applied to the activations at runtime, and mlx-lm's plain loader would decode plausible garbage without raising, so chad routes the pack to its own loader. The decode fast-path fuses gate|up, qkv|z and q|k|v behind one rotation each and folds the pack's sign vectors into the weights at install, so a rotation is one kernel in the activation dtype on every path (serial step, verify forward, prefill) where the pack's own is four; the small-M verify kernel now covers 2-bit g128 (mlx's stock 2-bit matmul made an 8-wide verify cost 9× a step and drafting a net loss) and tiles widths 9-24 into 8-row calls, because the same cliff sat one row past the tile (a 9-token forward: 266 ms → 147; 16 tokens: 343 → 203); the speculative schedule is seeded with the round-cost ladder measured on these weights instead of the 3-bit's, which prices every width past one 9-18% too cheap here, and each turn now starts from the ladder the last one measured (not a measured tok/s change: on low-acceptance text the schedule sits at break-even with serial on either seed); a checkpoint of the shipped shape with no bundled drafter borrows the shipped one, so pointing --model at the upstream prism-ml pack works too. The pack's chat template defaults reasoning_effort to xhigh; chad passes medium unless CHAD_REASONING_EFFORT is set, on every render including the cross-project warm-prefix head (rendered without it, the head was not a prefix of the real prompt under the upstream template and every fresh directory paid the cold prefill). CHAD_PRISM_ROT_FP32 is the A/B arm for the rotation's precision. A Prism pack the fast-path cannot fuse now says so, with the reason, instead of decoding several times slower in silence; sign vectors are shared by content rather than by first-seen-per-width, which could trigger exactly that; and a donor drafter whose config and weights resolve to different hub snapshots is healed or refused rather than handed to the loader half-empty.