Hash Verification Confirms OrcaRouter's Qwen3.8-Flash-Next-Uncensored-GGUF Uses Targeted Abliteration

A file-by-file hash comparison of the OrcaRouter Qwen3.8-Flash-Next-Uncensored-GGUF upload against Qwen's original weights reveals that 80 of 131 weight files are byte-for-byte identical. The remaining 51 differing files contain matrices that the model card identifies as edited, confirming the use of abliteration (targeted editing) rather than a full fine-tune or model merge.

Verification Methodology

The analysis utilized Hugging Face's SHA256 hashes to compare the uploader's full-precision copy with Qwen's original at pinned revisions. Key findings include:

• 51 files differ in content but maintain identical sizes.
• Each differing file holds at least one of the 149 matrices listed as edited in the model card.
• None of the 80 unchanged files contain any of the listed edited matrices.
• 173 tensors, representing 72.6% of the model's file bytes, are provably untouched.

Performance and Limitations

While the hash verification confirms the structural changes, it does not reveal the specific nature of the edits. The uploader reports that on XSTest's 250 harmless prompts, refusal rates dropped from 9.6% to 1.2% with thinking disabled. Capability scores generally remained within 2.3 points of the original, except for a 2.3-point drop on 300 MMLU questions. These figures are unreproduced by the analyst, who did not run the model.

Metadata Discrepancy

An anomaly was detected in the upload's metadata: it names Qwen as the direct base model, while the model card names the edited parent. This inconsistency causes Hugging Face's model tree to skip a generation in the lineage display.