Open-Weight Uncensored Gemma 4 31B Lands on Hugging Face
An uncensored, open-weight modification of Google’s Gemma 4 31B instruction-tuned model was published on Hugging Face on 5 November 2026, and compressed files for local use followed two days later.
The repository, richardyoung/gemma-4-31B-it-ARA-heretic, describes the model as a “decensored” derivative of google/gemma-4-31B-it, produced with Heretic 2.0’s Arbitrary-Rank Ablation method. The Hugging Face API timestamps the upload at 15:11 UTC on 5 November (18:11 in Europe/Istanbul). This kind of ablation does not retrain the model from scratch. It locates internal directions associated with refusals and dampens them, layer by layer.
The model card reports that, on a 100-prompt development set, refusals fell from 99 for the original model to 28, with a KL divergence of 0.0845 from the base model’s answers on harmless prompts. A held-out set of 972 prompts from WildGuard is reported as falling from an 89.3 percent refusal rate to 0.6 percent. Average performance across five capability benchmarks is reported as 2.18 percentage points lower than the original.
Those figures come from the author, from a single optimisation run, with refusal labels assigned by automatic judges and no human audit, as the card itself notes. Residual refusals remain. The card also states that the residual measures willingness to answer under greedy decoding, not how harmful an answer is.
On 7 November at 05:50 UTC (08:50 in Europe/Istanbul), the quantisation account mradermacher published static GGUF conversions. The card flags a Q4_K_M file of 18.8 GB as fast and recommended, and includes optional multimodal projection files so the 31B build can still take image input. As of 13 November, Hugging Face showed 96,342 downloads in the past month on the GGUF repository, compared with 435 downloads on the full-weight original.
The licence is Apache 2.0, inherited from Gemma 4, with the use-based restrictions of OpenRAIL++ added on top. The card says the base licence prevails in a conflict, that the model’s refusal behaviour was reduced for research on abliteration, and that it must not be used to generate content for any prohibited use listed in the OpenRAIL++ text. A lower refusal rate is not permission to ignore that licence, or the law.
For people who run uncensored models locally, the release matters because Gemma 4 31B is a relatively strong, multimodal open model, and a build that answers more of the prompts a stock instruction model turns down can be run on a 24 GB card at the recommended quant. It is not a dedicated roleplay fine-tune, and the author’s own metrics show both leftover refusals and a small cost in benchmark scores. Anyone using it remains responsible for what they generate and for the terms that travel with the weights.