GGUF quantizations of openPangu-2.0-Flash, Huawei's 92B-A6B MoE trained to 512K context, converted for ik_llama.cpp.

The OpenPangu architecture merged into ik_llama.cpp through PR #2065. These are self-contained conversions with the DSA/SWA schedule and indexer configuration stored in GGUF metadata.

The IQ4_NL file is 3.25 GiB, or 5.7%, smaller than the Q4_K_M control.

The importance matrix was collected from 100 independent 512-token chunks drawn from mixed Chinese and English documentation, code, and general prose. Its 700 entries cover all 44 base MoE blocks. The standalone collector does not exercise the three NextN/MTP blocks, so blocks 46 through 48 use the reviewed Q4_K_M tensor policy instead of uncalibrated importance data.

The final GGUF contains 1,681 tensors:

MoME convolution weights, mHC tensors, attention sinks, and DSA indexer weights remain at their reviewed higher-precision types. Absorbed MLA K-B tensors use Q8_0 and V-B tensors use Q4_K. The quantization source was the BF16 conversion, not the existing Q4_K_M file.

SHA-256:

CPU:

Basic CUDA placement, with routed experts in system RAM:

--jinja is required because the embedded chat template carries the model's message format. The template enables thinking under the kwarg name thinking; pass chat_template_kwargs: {"thinking": false} in a request to disable it.

NextN/MTP speculation is optional, up to heads=3.

The Q4_K_M file uses --custom-q overrides for convolution, mHC, sink, and indexer tensors. The k=3 MoME convolution tensors are not block-quantization compatible, while the other preserved families cost little relative to the full model.

Its SHA-256 is:

This is a conversion of Huawei's openPangu-2.0-Flash weights and inherits the openPangu License, which includes an attribution ("Powered by openPangu") notice requirement and a European Union use restriction. Read the license before deploying.

Powered by openPangu.