Qwen3.8-27B-GeoRefine-D14-abl-MIXED-IQ4_XS
Parent model: huihui-ai/Huihui-Qwen3.8-27B-abliterated
Parent revision: 739e3c5b89849f6c238ce1e5b70008612ae42cdd
Parent license: apache-2.0
Candidate status: candidate
Certification status: uncertified
Primary weight SHA-256: afe9e0cb33a232fe4735635e9cbb95a2f3fa4d8de2990de0d5fdb9e8a979299d
This repository packages an existing, lossy mixed-precision quantization of huihui-ai/Huihui-Qwen3.8-27B-abliterated, pinned to commit 739e3c5b89849f6c238ce1e5b70008612ae42cdd. It is experimental and uncertified. Complete behavioral and runtime certification are not established.
Artifact identity
| Field | Value |
|---|---|
| Canonical file | Qwen3.8-27B-GeoRefine-D14-abl-MIXED-IQ4_XS.gguf |
| Source file | D14-abl.gguf |
| Size | 14,199,740,480 bytes |
| SHA-256 | afe9e0cb33a232fe4735635e9cbb95a2f3fa4d8de2990de0d5fdb9e8a979299d |
| Identity evidence | Fresh full-file SHA-256 recorded by the read-only mesh identity receipt on old-donkey |
| Tensor descriptors | 866 tensors; GGUF v3; 51 metadata entries; 32-byte alignment |
| Most frequent quantized type by tensor count | IQ4_XS (F32 tensors excluded from this designation) |
Type counts from the GGUF header: F32 360, IQ4_XS 292, IQ3_S 76, Q8_0 128, Q4_K 9, Q5_K 1. The model uses a mixed tensor layout; the filename names the most frequent quantized type by tensor count and does not imply a uniform quantization.
Provenance and limits
- Parent:
huihui-ai/Huihui-Qwen3.8-27B-abliterated@739e3c5b89849f6c238ce1e5b70008612ae42cdd. The smallconfig.json,generation_config.json,tokenizer_config.json, andchat_template.jinjafiles are pinned sidecars from that revision. GGUF tokenizer metadata is embedded in the weight artifact; large upstream tokenizer vocab files are not bundled here. - Transformation: Mixed precision as encoded in the source GGUF; exact tensor types and counts are reported in this manifest.
- Calibration: The historical stock Qwen3.8-27B imatrix SHA-256 0405bcb37a6ab0b7bb94ccb98d0fb4d8e87fd53daf9750eb4bb331aceabbc7a3 was reused for the abliterated parent in v1; fresh parent imatrix is v2. This artifact uses the v1 stock-matrix calibration.
- This D14-abl mesh artifact is distinct from the D14 research artifact in the portfolio registry. D14 research grades and its historical SAE failure do not transfer to this file.
- No complete capability evaluation is established. This artifact has no quality verdicts attached; D14 research results are from a distinct artifact and are not inherited.
- Quality gates: artifact identity PASS (source-host full-file digest); behavioral quality UNTESTED; SAE certification UNTESTED; runtime qualification UNTESTED. No behavioral PASS is claimed.
- The header-only inspection read tensor names, shapes, types, and offsets, never tensor payload bytes. The supplied identity receipt verified the full artifact separately.
Use python3 verify_artifact.py next to the GGUF to check exact byte count and SHA-256 with an 8 MiB streaming buffer. Download the GGUF and verifier from this repository into the same directory before running the check.
See artifact-manifest.json for machine-readable identity, tensor type counts, parent pin, sidecar digests, status, and limitations. NOTICE records the upstream attribution and declared license.