[{"data":1,"prerenderedAt":38},["ShallowReactive",2],{"profile-gemma-4-26b-a4b-uncensored":3},{"slug":4,"title":5,"seoTitle":6,"description":7,"summary":8,"published":9,"updated":9,"reviewed":9,"html":10,"toc":11,"sources":25},"gemma-4-26b-a4b-uncensored","Gemma 4 26B A4B uncensored and abliterated variants","Gemma 4 26B A4B Uncensored: Heretic and Abliterated","Gemma 4 26B A4B uncensored variants: abliterated, QAT-based and Heretic builds with reported refusal and KL figures, licenses, hosting and memory needs.","Gemma 4 26B A4B is Google DeepMind's mixture-of-experts model with 25.2B total and 3.8B active parameters, 256K context and image input, released under Apache 2.0. Its refusal-reduced variants include huihui-ai's abliterated weights, an abliterated QAT build, two Heretic builds that publish refusal and KL figures, and a HauhauCS GGUF with an undisclosed method.","2026-11-14","\u003Ch2 id=\"what-the-base-model-is\">What the base model is\u003C\u002Fh2>\n\u003Cp>Gemma 4 26B A4B is the medium mixture-of-experts model in Google DeepMind’s Gemma 4 family. The “A4B” means about 4 billion parameters are active per token; the card gives 25.2B total and 3.8B active, with 128 experts plus one shared expert and 8 used per token. It takes text and images and supports a configurable thinking mode.\u003C\u002Fp>\n\u003Cp>Two details matter for running it. First, low active parameters make it fast per token, but every expert must sit in memory, so it needs about as much memory as a 25B dense model. Second, its attention is hybrid: 25 of its 30 layers use a sliding window of 1,024 tokens and only 5 keep the full context. That keeps the KV cache small even at long context, which the \u003Ca href=\"\u002Ftools\u002Fvram-estimator\u002F\">VRAM estimator\u003C\u002Fa> models from the configuration file.\u003C\u002Fp>\n\u003Cp>Google publishes Gemma 4 under the Apache License 2.0. Earlier Gemma generations used the Gemma Terms of Use, and some variant repos still carry that older tag; the base license is the one that governs the weights.\u003C\u002Fp>\n\u003Ch2 id=\"the-variants-at-a-glance\">The variants at a glance\u003C\u002Fh2>\n\u003Cdiv class=\"table-wrap\" role=\"region\" aria-label=\"Scrollable table\" tabindex=\"0\">\n\u003Ctable>\n\u003Cthead>\n\u003Ctr>\n\u003Cth>Variant\u003C\u002Fth>\n\u003Cth>Method (per card)\u003C\u002Fth>\n\u003Cth>What the card reports\u003C\u002Fth>\n\u003C\u002Ftr>\n\u003C\u002Fthead>\n\u003Ctbody>\n\u003Ctr>\n\u003Ctd>huihui-ai abliterated\u003C\u002Ftd>\n\u003Ctd>Abliteration\u003C\u002Ftd>\n\u003Ctd>First 5 layers and the experts not edited; may still add warnings but not refuse\u003C\u002Ftd>\n\u003C\u002Ftr>\n\u003Ctr>\n\u003Ctd>huihui-ai QAT abliterated (GGUF)\u003C\u002Ftd>\n\u003Ctd>Abliteration of Google’s quantization-aware checkpoint\u003C\u002Ftd>\n\u003Ctd>No measurements\u003C\u002Ftd>\n\u003C\u002Ftr>\n\u003Ctr>\n\u003Ctd>coder3101 heretic\u003C\u002Ftd>\n\u003Ctd>Heretic 1.2.0, ARA method\u003C\u002Ftd>\n\u003Ctd>11\u002F100 refusals vs 100\u002F100; KL divergence 0.0499\u003C\u002Ftd>\n\u003C\u002Ftr>\n\u003Ctr>\n\u003Ctd>llmfan46 ultra-uncensored heretic\u003C\u002Ftd>\n\u003Ctd>Heretic 1.2.0, ARA method\u003C\u002Ftd>\n\u003Ctd>3\u002F100 refusals vs 100\u002F100; KL divergence 0.1237\u003C\u002Ftd>\n\u003C\u002Ftr>\n\u003Ctr>\n\u003Ctd>HauhauCS QAT Balanced (GGUF)\u003C\u002Ftd>\n\u003Ctd>Not disclosed\u003C\u002Ftd>\n\u003Ctd>“0\u002F465 refusals” on an unpublished test; MTP draft head\u003C\u002Ftd>\n\u003C\u002Ftr>\n\u003C\u002Ftbody>\n\u003C\u002Ftable>\n\u003C\u002Fdiv>\n\u003Cp>The two Heretic builds show the usual tradeoff. The llmfan46 build reports fewer refusals than coder3101’s but a KL divergence about two and a half times higher, meaning its predictions moved further from the original. Both numbers come from Heretic’s default prompt set as run by each creator, so treat them as a comparison between these two builds of the same base, not as a quality score.\u003C\u002Fp>\n\u003Cp>Quantization-aware training (QAT) means Google trained a checkpoint to lose less quality when compressed to 4 bits. Two variants start from that checkpoint, which is why HauhauCS ships only a Q4_K_M file and says higher-precision files would add size with little gain.\u003C\u002Fp>\n\u003Ch2 id=\"which-one-for-which-need\">Which one for which need\u003C\u002Fh2>\n\u003Cul>\n\u003Cli>\u003Cstrong>For published numbers and two hosting options:\u003C\u002Fstrong> coder3101’s Heretic build. Venice serves it in a trusted execution environment with end-to-end encryption, according to its model list, with a 64,000-token context and 4,096 tokens of output, but lists neither image input nor function calling for that offering. Featherless serves the same repo with a 32,768-token context.\u003C\u002Fli>\n\u003Cli>\u003Cstrong>To run at 4-bit on one consumer GPU:\u003C\u002Fstrong> one of the QAT-based GGUF files. Weights take roughly 15 GiB at Q4_K_M, plus a 1.2 GB vision projector if you need image input.\u003C\u002Fli>\n\u003Cli>\u003Cstrong>For the most documentation on what changed:\u003C\u002Fstrong> huihui-ai’s release, which says which layers and modules it left alone.\u003C\u002Fli>\n\u003C\u002Ful>\n\u003Cp>MolochAI lists \u003Ccode>gemma4-26b-a4b-qat\u003C\u002Fcode>, described as a refusal-reduced QAT build, without naming the weights. Rates for all of these are below.\u003C\u002Fp>\n\u003Ch2 id=\"limits-and-open-questions\">Limits and open questions\u003C\u002Fh2>\n\u003Cp>The huihui-ai card says it left “the 256 experts per layer” unedited, while Google’s card lists 128 experts plus one shared expert. It is probably a typo, but it is the kind of mismatch to note when a card is your only documentation. The HauhauCS repo is tagged with the Gemma Terms of Use license although Gemma 4 is Apache 2.0.\u003C\u002Fp>\n\u003Cp>No variant publishes a quality benchmark after the edit. Before relying on image input, test it: editing the language layers can change how the model describes what it sees.\u003C\u002Fp>\n",[12,16,19,22],{"id":13,"text":14,"level":15},"what-the-base-model-is","What the base model is",2,{"id":17,"text":18,"level":15},"the-variants-at-a-glance","The variants at a glance",{"id":20,"text":21,"level":15},"which-one-for-which-need","Which one for which need",{"id":23,"text":24,"level":15},"limits-and-open-questions","Limits and open questions",[26,30,34],{"title":27,"publisher":28,"url":29},"google\u002Fgemma-4-26B-A4B-it model card","Hugging Face (Google)","https:\u002F\u002Fhuggingface.co\u002Fgoogle\u002Fgemma-4-26B-A4B-it",{"title":31,"publisher":32,"url":33},"Gemma 4 license (Apache License 2.0)","Google AI for Developers","https:\u002F\u002Fai.google.dev\u002Fgemma\u002Fdocs\u002Fgemma_4_license",{"title":35,"publisher":36,"url":37},"Heretic","GitHub (p-e-w)","https:\u002F\u002Fgithub.com\u002Fp-e-w\u002Fheretic",1794622336923]