An abliteration that kept pace with the release

Zhipu's upstream repository appeared on 7 December 2025, and the Z.ai release blog followed on 8 December. The family has two sizes: GLM-4.6V at 106B for clusters, and GLM-4.6V-Flash, marketed as 9B, aimed squarely at local deployment. The Hugging Face model API counts 10,292,777,472 BF16 parameters in the exact artifact — vision encoder included — which matches the upstream count to the parameter. Huihui's repository was created at 17:00 UTC on 9 December; the upstream repo's last touch landed about an hour later. The abliteration happened essentially at release speed.

The architecture is unchanged: Glm4vForConditionalGeneration, image-text-to-text, the same roughly 20.6 GB of weights in five shards. What changed is a subset of the text-side weights. The model card is explicit: “It was only the text part that was processed, not the image part.”

Text-only abliteration, and what that really means

Abliteration is the refusal-direction surgery documented by Arditi and colleagues and implemented for this workflow by Sumandora's remove-refusals-with-transformers, which the card links. It estimates a direction in activation space associated with refusal and projects it out of selected weights. It is a weight edit, not a jailbreak prompt and not fine-tuning.

The caveat is in that one sentence of the card. Refusal in a vision model can be conditioned on what the model sees — a violent image, a medical scan, a product label — and the refusal machinery that reads image features lives partly outside the text pathway. Huihui only touched the text part. So “uncensored vision model” is really “an uncensored language model that can see.” Whether image-conditioned refusals survive the edit is an open question, because neither huihui nor anyone else has published an evaluation of this derivative: no refusal-rate numbers, no post-edit rerun of vision benchmarks. The only post-release corroboration I found is community chatter — a July 2026 uncensored-model roundup repeats the text-only point verbatim.

Everything Zhipu claims for the base model — SoTA-in-class visual understanding across MMBench, MathVista and OCRBench, 128K training context, native multimodal function calling — is an upstream claim. Zhipu also self-admits real limits that carry straight over: pure-text QA is comparatively weak, the model can overthink or repeat, and perception (counting, identifying individuals) is unreliable. None of that is fixed by an abliteration.

What is proven here?

The exact artifact, license and architecture are inspectable, and the edit is documented as text-only. Upstream capability claims belong to Zhipu's checkpoint. Huihui publishes no refusal-rate or capability-retention evaluation, and no published study covers image-conditioned refusal on this derivative.

Why the MIT part is the rare part

Abliterations inherit the upstream license, and most flagship models that get the treatment ship Apache-2.0 or something stricter. MIT for a genuinely usable open vision model is the exception — Zhipu chose it for Flash, huihui kept it, and downstream quantizers are free to do MIT-permitted things with it. One of them already did: the AliBilge GGUF port carries an Apache-2.0 tag, a relicensing MIT permits. For commercial local use, that is the cleanest legal envelope in the uncensored-model space: no copyleft, no fee, attribution only.

The sleeper framing is real. The exact artifact has 99 downloads and 18 likes; its own quantized shadow, seanbailey518's GGUF from 30 December, accumulated 2,051 downloads. The audience found the quants, not the source.

The first quant wave stalled at the runtime

December 2025 through January 2026 produced a respectable quant ecosystem: the AliBilge Ollama package (which has since logged over 9,000 pulls), seanbailey518's GGUF, and a full q2-to-q8 MLX set for Apple Silicon from January. And yet the only discussion thread on the model is someone asking for Ollama support, with a reply from the publisher that aged badly: “Ollama probably doesn't support all the functionalities completely yet. We have tested it, and it crashes when trying to recognize images.” The follow-up in January pointed at ollama issue #13391 and went nowhere. For months, running this model in llama.cpp-style toolchains meant running it blind — the exact modality an abliteration of a vision model shouldn't be reduced to.

What actually changed in August

On 17 August 2026, mradermacher published two repos: a standard GGUF ladder with an f16, and an i1-imatrix ladder calibrated with an imatrix file. The telling detail is in the first one's file list: .mmproj-f16.gguf and .mmproj-Q8_0.gguf — the multimodal projector, the piece that lets llama.cpp-family runtimes actually see. GGUF metadata carries a 131,072-token context. Both repos point their base_model tag directly at huihui-ai's artifact, and both had zero downloads when I checked.

So this is a supply-side signal, not a demand spike: someone with a conversion pipeline decided the GLM-4V runtime support had matured enough to ship vision-capable quants. It quietly never left the local-model conversation either — an April 2026 local-models roundup still listed a Huihui-GLM-4.6V-Flash-abliterated-GGUF tag. The model was dormant, not dead.

How we treat it

Our prepared profile pins revision 7d7926ee0a8d02e46bdab97ead8d5396bfc071df, serves it on a single L40S at an approximate managed price estimate of $2.34/hour, and treats the edit as text-only until someone measures otherwise. We do not transfer Zhipu's vision benchmarks to this endpoint, and the publisher's own warnings — significantly reduced safety filtering, research and controlled environments only — apply in full. A permissive-license, refusal-reduced multimodal checkpoint is a useful research and tooling artifact. That is what this is.

The idea, in plain words

Why an MIT license matters more than the benchmark table

MIT is the 'do whatever you want' license: commercial use, modification, redistribution, no strings attached. Most capable vision models are not MIT. For a company, license terms decide whether a model is usable at all — which is why a smaller MIT model can beat a bigger restricted one in practice.

Primary sources