Gemma-4 26B abliterated checkpoints land on HuggingFace with multimodal option

Two abliterated Gemma-4 26B checkpoints surfaced on HuggingFace on May 10–11, each targeting practitioners running uncensored models locally. Netr0y's gemma-4-26B-A4B-it-ultra-uncensored-heretic-nvfp4 is an 8-bit compressed-tensors quantization built on llmfan46's base heretic fine-tune, while gsting's gemma-4-26B-A4B-it-uncensored-heretic ships as a full-precision image-text-to-text pipeline with abliterated safety layers. Both carry the "heretic" label—shorthand in the uncensored scene for models with alignment guardrails surgically removed.

The Netr0y build has logged 50 downloads since posting; gsting's version sits at zero downloads but picked up a like within hours. Neither card includes benchmark numbers, sample outputs, or quantization details beyond the safetensors tag, so VRAM requirements and actual refusal behavior remain unverified. The image-text-to-text tag on gsting's checkpoint suggests multimodal capability, though the model card doesn't specify whether that's native Gemma-4 vision support or a bolt-on adapter. The compressed-tensors format on Netr0y's quant should fit consumer GPUs with 24GB VRAM, assuming the 8-bit claim holds, but no inference logs or perplexity scores have been posted to confirm quality loss from quantization.

Gemma-4 26B is Google's latest open-weight release, positioned between the 9B and 70B tiers and shipped with instruction tuning and safety filters intact. These heretic variants reverse that work. What remains to be tested: runtime benchmarks, refusal-rate comparisons against the base Gemma-4 26B-it checkpoint, and whether the image-text pipeline actually works end-to-end or requires separate vision preprocessing. If either uploader posts eval numbers or sample prompts in the coming days, the community will know whether these builds hold up under load or collapse into incoherence at high compression.