2.3.0

SORRRY The ternary default was a mistake; the default is now Unsloth's UD-Q3_K_XL GGUF

We shipped the wrong weights. The ternary pack bought a big context window, and the tasks we tested it on were too easy to show what it cost. On 36 paired polyglot tasks it passes 24 where Unsloth's Qwen3.8-27B-UD-Q3_K_XL.gguf passes 34 (10 up, 0 down, p = 0.002), and the GGUF gets there in 0.30× the generated tokens and 0.38× the wall clock. All twelve ternary failures ran out the 20-minute wall clock. Design has the account; every run cited is in benchmarks/polyglot/RUNS.md.

chad now reads GGUF files natively. Converting Unsloth's file to MLX's affine format passes 6 of 9 trials where the file passes 9: its i-quants have no MLX equivalent. Serving it from llama.cpp would give up the engine-owned cache. So the llama.cpp blocks load as they are, and chad's Metal kernels decode them (gguf_pack.py, mlx_gguf.py; 15 formats, bit-exact to llama.cpp, MIT port credited in NOTICE).

Re-download: ~14 GB, up from ~8. The GGUF plus a 1.2 GB sidecar (nathansutton/Qwen3.8-27B-DFlash2-MLX) with the tokenizer and drafter. Free the old snapshot with hf cache rm.
Less context: ~74k on 24 GB, down from ~150k. The median trial peaked at 8.6k. --model takes any Unsloth file of this model, as a local .gguf path or owner/repo/file.gguf: UD-IQ3_XXS (10.9 GB) gets ~138k and passed 8 of 9. The ternary and affine packs still load.
Serial decode is ~12 tok/s, level with llama.cpp on the same file; drafted, ~46. The affine repacks decoded faster serially (18-21 tok/s) and lost on the tasks. A real agent session decodes at a ~21 tok/s median.
First start ~75 s, later starts ~9 s. The first load saves its converted arrays to ~/.cache/chad/gguf/ (~13 GB more disk, counted by the preflight).