Mero-Artemis-31B-v0.3.1-heretic: uncensored 31B multimodal merge on Gemma4
sh0ck0r released Mero-Artemis-31B-v0.3.1-heretic, a 31-billion-parameter uncensored image-text-to-text model built on Gemma4 and available in safetensors format on HuggingFace.
Mero-Artemis-31B-v0.3.1-heretic is a 31-billion-parameter image-text-to-text model from sh0ck0r, released this week on HuggingFace. The checkpoint is a mergekit-built fusion using Gemma4 as the base architecture, tagged "not-for-all-audiences" and "heretic" — explicit markers that safety guardrails have been removed. Weights ship in safetensors format.
The model card carries zero downloads and zero likes at publication time, typical for a fresh upload. No benchmark numbers, sample outputs, or merge recipe appear in the listing yet. The image-text-to-text pipeline tag indicates multimodal capability — text and image inputs, text output — though the card does not specify context length, VRAM requirements, or which Gemma4 variant was used as the donor.
What stands out
- 0131B parameters — midsize by 2025 standards, likely runnable on a single 48GB or dual 24GB consumer setup with quantization.
- 02Gemma4 base — Google's Gemma4 series supports native multimodal input; this merge inherits that architecture.
- 03Mergekit fusion — the checkpoint is a model merge, not a from-scratch train, meaning it blends weights from multiple sources (recipe not disclosed).
- 04Uncensored tags — "not-for-all-audiences" and "heretic" markers signal the model is intended for unrestricted use without safety enforcement.
- 05Safetensors format — modern, fast-loading weight format compatible with transformers, ComfyUI, and most inference stacks.
