Checked 2026-10-05. Huihui-ai publishes abliterated checkpoints and quantized variants on Hugging Face. Treat the publisher, the procedure, and each exported artifact as separate things: there is no single version-independent "huihui method" established by the inspected cards.
The machine-readable snapshot pins eight model revisions, their declared base models, card hashes, creation/modification dates and source URLs. These are publisher records, not local performance certifications.
Concrete disclosures worth knowing#
| Checkpoint / card | What the publisher reports | How to interpret it |
|---|---|---|
| Qwen3.8-27B GGUF, pinned card | Several newer variants list layers 22–52; an older UD section lists 17–52. The card also retains earlier notes and says MTP/vision were not modified. | Ranges are zero-based and variant-specific. Record the exact file and section; do not collapse the whole repository into one layer recipe. |
| GLM-5.3-Flash GGUF, pinned card | Layers 15–35 are edited; other layers and expert modules are described as untouched. The card names an upstream quantized source. | This is a selective-support claim, not proof that expert preservation guarantees behavioural/capability preservation. |
| Qwen3-4B v2 and 14B v2 | Their notes associate a layer/candidate-layer change with correcting garbled output. | Useful motivation for layer ablations and coherence checks. It is not a universal instruction to exclude layer zero. |
| Qwen3-30B-A3B Instruct | Describes a faster, improved approach while linking the Transformers-based reference implementation. | The inspected card does not supply a complete benchmarked recipe proving the comparative claim. |
| Gemma-4-31B | Claims modification of both thinking and non-thinking modes. | Evaluate both modes separately with the correct chat template; the claim is not itself a measured score. |
The Qwen3.8-27B card also describes mixed tensor types: selected tensors use Q8_0 in some K_L variants and BF16 in Q8_0_L. File names therefore do not reliably predict uniform bit width or memory use. Its Ternary section requires a specific llama.cpp fork. These are card-reported packaging/runtime constraints; they were not tested here.
Recent entries are not necessarily new techniques#
The inspected listing includes Qwen3.8-Flash-Next, Xing4.0-29B-A4B, and the GLM variant above. The first two were created on September 29; the GLM repository on September 26. Recent modification timestamps can describe card edits, metadata fixes or replacement shards rather than new model training.
Check per-model licensing and dependencies. For example, the Flash-Next card declares a Qwen community license, so the publisher name must not be used to infer one shared license for every model.
Reference implementation and inspection tools#
Many cards link Sumandora/remove-refusals-with-transformers at the reviewed commit. That establishes a cited Transformers-based reference, not proof that each huihui export was generated using exactly that revision and defaults.
huihui-support/compare_models computes mean absolute weight differences and plots them. The reviewed script accepts --model1 and --model2. A heatmap can help locate changed matrices, but cannot recover the original contrast data, prove a particular edit algorithm, or establish retained capability. The tooling record preserves source hashes.
Evaluate a candidate without confusing causes#
- Pin the base model and candidate revision, tokenizer/processor, template, inference runtime, quantization and decoding settings.
- Compare like-for-like artifacts. A quantizer, template or prompt-format change can mimic an abliteration effect.
- Test task quality, false refusals, coherence, languages, thinking modes and tool use separately.
- For a selective-edit claim, inspect tensor names/shapes and record changed versus unselected tensors. Weight similarity is not behavioural equivalence.
- Report repeated measurements where sampling matters, together with uncertainty and failure cases.
- Keep an original checkpoint and a rollback path; apply the existing evaluation gates before treating a community release as suitable for a deployment.
The source snapshot records exactly what was checked. It does not rank these checkpoints or claim the latest upload is the best choice.