Lambent releases Arsenic-Shahrazad 12B v4.3, uncensored Mistral fine-tune

Lambent released Arsenic-Shahrazad 12B v4.3 this week, an uncensored 12-billion-parameter fine-tune built on the Mistral architecture and hosted on HuggingFace under a CC-BY-NC-4.0 license. The model carries HuggingFace's "not-for-all-audiences" tag and is distributed in safetensors format, the standard tensor serialization used across ComfyUI, text-generation-webui, and llama.cpp.

The 12-billion-parameter scale is designed for consumer hardware—a single RTX 4090 or two RTX 3090s can run the model at full context, making it accessible to hobbyists and small studios without enterprise GPU clusters. Mistral's efficient attention mechanism and strong performance-per-parameter have made it a popular base for fine-tuning work in the uncensored space.

The CC-BY-NC-4.0 license allows modification and redistribution for non-commercial use but blocks commercial deployment without separate licensing. This pattern has become standard among open-weight creators balancing openness with revenue protection. Developers building commercial SaaS or APIs will need to negotiate directly with Lambent or select fully permissive alternatives like Apache 2.0 or MIT-licensed models. Arsenic-Shahrazad v4.3 is a fine-tune of Lambent's v4.1 checkpoint, released earlier this year, though Lambent has not published a detailed changelog documenting the specific improvements.