Niwaki (庭木): every routed expert individually width-pruned to the neurons its own routed tokens actually use, reconstructed to compensate, then briefly distilled from the full model, and stored at low precision. A paper with the full method is coming soon.

Reference Qwen3.6-35B-A3B at Q8_0 measures 6.95 under the identical protocol (llama-perplexity, WikiText-2 test, 512-token windows). These llama.cpp numbers are not directly comparable to the MLX repo's 2048-window benchmarks; the relative standings match across both.

Generation battery (measured on the canonical MLX weights; reference scores 0.89 / 0.78): bigram-diversity avg/min = 0.82 / 0.62 across an 8-prompt code/reasoning/chat/creative battery — see the limitation note below.

The recommended UD-Q3K build is quantized structure-aware (importance matrices calibrated on the same code-weighted corpus as the model itself), mirroring the artifact's native allocation: the always-active backbone (attention, shared experts, embeddings) is kept at high precision (Q6_K) while the pruned routed experts ride a compact carrier (q3_k, imatrix-guided). It matches or beats uniform Q4_K_M quality at ~20% fewer bytes on this model family.

Known limitation (measured on the MLX canonical): sustained code generation is the weakest axis of this model (battery minimum 0.62 on long-file code prompts, vs 0.75+ for its larger siblings); output tends to truncate early rather than stay coherent to the end of a long file. Chat, reasoning, and short-form writing are functional. Choose the 19B for heavy code work.

Requires a recent llama.cpp with Qwen3.6 (hybrid linear-attention) support. Canonical benchmarks, method outline, and the MLX-native artifact: Qwen3.6-11B-A3B-Niwaki-4bit-mlx. Family: 27B-2bit · 19B-2bit.

Base model by the Qwen team (Apache 2.0); pruning, distillation, and GGUF builds by the Niwaki project, 2026-08.