Niwaki (庭木) are Japan's garden trees, sculpted by meticulous pruning so that every branch serves the form of the whole. This model applies that spirit to a Mixture-of-Experts. A paper with the full method is coming soon.

First-generation GGUF builds under the identical protocol:

This build beats even the first-generation flagship (27B, 10.41) at 1.3 GB fewer bytes.

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 0.63 / 0.51 under the identical battery): bigram-diversity avg/min = 0.80 / 0.56 across an 8-prompt code/reasoning/chat/creative battery.

The recommended UD-Q3K build is quantized structure-aware (importance matrices calibrated on a mixed corpus), mirroring the artifact's native allocation: the always-active backbone (attention, shared experts, embeddings) is kept at high precision (Q6_K) while the routed experts ride a compact carrier (q3_k, imatrix-guided). It stays within 1.5% of uniform Q4_K_M quality at ~18% fewer bytes; pick Q4_K_M if you want the last percent of quality over size.

Format note: GGUF requires a uniform expert count per model, so the shared-only layers carry zero-valued expert tensors stored at ~1.6 bits/weight (~3.1 GB of the file). They contribute nothing to outputs; this is why these files are larger than the MLX repo at equal quality.

Requires a recent llama.cpp with Qwen3.6 (hybrid linear-attention) support. Canonical benchmarks and the MLX-native artifact: Qwen3.6-19B-A3B-Niwaki-v2-4bit-mlx. Family: v2-2bit · first-generation 27B-2bit · 19B-2bit · 11B-4bit.

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