- Published on
A 9B model that refuses almost nothing is this week's most-liked download
- Authors

- Name
- AbnAsia.org
- @steven_n_t
Listen to this article · A 9B model that refuses almost nothing is this week's most-liked download
0:00 / 0:00
Watch this article · A 9B model that refuses almost nothing is this week's most-liked download
A community model published on 30 October has become one of the most-downloaded small models on Hugging Face: 27,037 downloads and 7,590 likes, for a 9B-parameter model whose entire pitch is that it will not refuse you.
xtools-at/DeepSeek-V4-Pro-Qwen3.5-9B-UNCENSORED-heretic is a decensored version of a reasoning fine-tune called Jackrong/DeepSeek-V4-Pro-Qwen3.5-9B. The numbers in its card are the reason to look at it: on the author's 100-prompt refusal test, the base model refuses 100 of 100. This one refuses 3.
What the model actually is
Two layers of work sit under that name, and only the second is the scandal.
The first is a distillation: DeepSeek-V4-Pro-Qwen3.5-9B takes Qwen3.5-9B and trains it on roughly 250,000 math and STEM samples generated by DeepSeek-V4-Pro, in what its author calls "Max Effect" mode. That is a normal, respectable piece of work — a small model that inherits the reasoning style of a much larger one, in six languages (English, Chinese, Korean, Japanese, Spanish, Russian), under an Apache-2.0 licence.
The second is the edit. xtools-at ran a tool called Heretic (v2.0.0.dev0), which performs what it calls Arbitrary-Rank Ablation on the model's own weights. The card publishes the exact parameters: layers 9 to 24, preserve_good_behavior_weight 0.8264, steer_bad_behavior_weight 0.0003, overcorrect_relative_weight 0.9645, 12 neighbours.
That is not a jailbreak prompt. It is a recipe.
DeepSeek-V4-Pro
A frontier model, used only as a teacher: its reasoning traces are generated in "Max Effect" mode.
~250K samples
Math and STEM problems and answers, distilled into a much smaller student.
Qwen3.5-9B
The student: nine billion parameters, six languages, Apache-2.0 — runnable on one GPU.
Heretic, layers 9–24
Arbitrary-Rank Ablation rewrites the middle layers; the card publishes every weight used.
The number that matters is not the refusal rate
Almost every headline about a model like this stops at "it answers anything". The interesting figure is the other one: KL divergence 0.0422.
KL divergence measures how far the edited model's output distribution has moved from the original. Zero would mean no change; a small number means the model still behaves like itself almost everywhere. The author's claim, in effect, is that you can delete the refusal behaviour while keeping the model.
Whether that holds is the open question, and the card is honest about the limits: it is the author's own test, on the author's own prompt set, with no independent evaluation. The base model's math claims — GSM8K 94.50% averaged over four runs on an NVFP4 checkpoint, 92.40% on a three-subset MMLU-Pro run — come from the same lineage and the same publication practice.
Why this is a governance story, not a gadget story
Abliteration is not new. What changed is that it is now a published, parameterised, repeatable operation on a model small enough to run on a workstation — and the result cleared 27,000 downloads in two weeks without a launch, a blog post or a vendor behind it.
That is the same point the Qwen4-Max release made from the other direction. There, the frontier arrived and the weights did not. Here, a nine-billion-parameter model arrived with its safety layer removed and its removal recipe attached.
For a company, the practical consequence is narrow and unglamorous: model provenance is now part of your supply chain. A downloaded checkpoint is not a product with a warranty. It is a file that someone has edited, and the edit is documented well enough that you can see what was done — which is more than you get from most fine-tunes.
What to do with it
If you run local models for extraction, classification, translation or red-teaming, this class of release is useful and cheap: 9B parameters, Apache-2.0, six languages, one GPU.
If you ship model output to customers, three rules survive this release unchanged. Run your own evaluation before you deploy a checkpoint you did not train. Treat a missing refusal as a missing test, not a feature. And keep the card: when something goes wrong, the parameters that were published are the closest thing to an audit trail the open ecosystem offers.
The model is not the story. The fact that its edits are reproducible by anyone with a GPU is.
Source: the model card and Hugging Face API entry for xtools-at/DeepSeek-V4-Pro-Qwen3.5-9B-UNCENSORED-heretic, read 16 November 2026 (27,037 downloads, 7,590 likes, created 30 October 2026, Apache-2.0); the linked base model Jackrong/DeepSeek-V4-Pro-Qwen3.5-9B and its benchmark tables; the refusal and KL figures are the author's own, as printed in the card. Heretic is named by the card as v2.0.0.dev0; we did not evaluate the tool or the resulting model. The analysis is ours.
Contact: [email protected]