Quantization: ZAYA1-8B is a MoE, 760M active per token
Quantization: picks by use case - ZAYA1-8B Q4_K_M for long/agentic coding on a small local model
Quantization: MiniMax-H3 is a video model, not an LLM the build failed to run
Home: status 2026-09-24 (HRX decode fix, prefill split, 1bit serve --mtp/--parallel/--adaptive/RAG, lean option, 1bit.gg, dx M3d); Quantization: UD sweet spots for the top models, sub-4-bit UD files run on HRX0
Quantization: the lean option, ROCmFP4 and ROCmI4 against UD-Q4_K_XL on Qwen3.8-27B
Quantization: Q4NX has three chunk kinds (q4_1, Q4_K, Q8), fully decoded; Q4_K repacks into q4_1
GPU page for the live ggml-hrx build (HRX0), re-measured at hrx-system 51b1739; Quantization table on HRX0; sub-4-bit UD fails on HRX0
Status: steps 1-3 merged, mlx, XDNA pin, ZINC + NVIDIA (RTX 5090); new NVIDIA and ZINC page; transformers GGUF reference
Wiki: NPU, Quantization, GPU, Apple, and working on the shared NPU
Measured results from 2026-09-22/23 with their sources: the XDNA2 int8 x int4 multiply mode
(4-bit is the NPU sweet spot; ternary converts to int4 exactly), Bonsai-8B in 4-bit on the NPU,
the open dx speed-up series (1bit-MONSTER #2675/#2676/#2680), how to read an ERT timeout,
Unsloth Dynamic GGUFs on the llama.cpp fork, Vulkan0 vs HRX20, lemon-mlx-engine and the Metal
backend on an Apple M4, and the rules for agents sharing the one NPU.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>