Qwen3-Coder 30B-A3B — MER Q4_0 Qualification Artifact
Private engineering artifact containing the exact Q4_0 files used to
qualify Micro-Expert-Router-SSD-Streamed-MoE (MER).
This is not yet a release-grade Amalgafy quantization or model-quality
benchmark. The GGUF was requantized from an existing quantized GGUF with
llama.cpp using --allow-requantize and --pure. Requantization may compound
quantization error.
Artifacts
- artifacts/gguf/Qwen3-Coder-30B-A3B-Instruct-pure-Q4_0.gguf
- artifacts/mer/qwen3-coder-30b-a3b-mer-q4_0-v1.tar.zst
- evidence/pr6-q4-parity.json
The MER archive contains 6,144 routed experts, 435 dense tensors, tokenizer,
configuration, metadata, and the canonical ggml-standard-v1 Q4_0 layout.
Qualification
Qualified on an NVIDIA L4 through WGPU/Vulkan using MER commit:
dac1d213cf641ba79a48e74c24f80bc2eca66548
Results:
- Seven raw WGSL Q4_0 cases passed
- Three complete checkpoint-expert vectors passed
- Initial expert installation occurred exactly once
- Subsequent vectors uploaded zero expert-weight bytes
- Zero CPU fallback or degraded expert execution
- Worst complete-expert absolute error: 7.6293945e-06
Checksums
- Pure Q4_0 GGUF:
8ddf61cadd354a5095905cc5ce535c44b777d0313ac241abcd2ceafa3362551b
- MER archive:
659b8d31d0a83292c632aa109c8edb5301f4041b1a60ef43c6f23ec0404061fe
- Parity report:
1d579a9e7ebc93191544ff162027e840dfbbd55ae7cc85e81021bb6e85784c60
Provenance
- Upstream: Qwen/Qwen3-Coder-30B-A3B-Instruct
- License: Apache-2.0
- llama.cpp: 030ebb558a5820b444a8f836ed5cdd46c9b4bd7a
- MER: dac1d213cf641ba79a48e74c24f80bc2eca66548
Qwen3-Coder is provided by the Qwen team. This repository preserves the
upstream license and identifies the conversion and requantization changes.
