Lambent releases Arsenic-Shahrazad 12B v4.1 under non-commercial license

Lambent released Arsenic-Shahrazad 12B v4.1 on HuggingFace this week, a 12-billion-parameter model built on the Mistral architecture. The weights are distributed under a CC-BY-NC-4.0 license, restricting commercial use while allowing modification and redistribution for non-commercial purposes. The model is tagged "not-for-all-audiences" on HuggingFace, signaling unrestricted output capability — a hallmark of the Arsenic-Shahrazad lineage.

At 12 billion parameters, the model should run on consumer GPUs with 24GB VRAM or larger, making it accessible to local deployment setups that prioritize uncensored generation. The v4.1 release is a fine-tune of Lambent's own Arsenic-Shahrazad 12B v4 base, suggesting iterative refinement rather than a ground-up retrain. The model card offers no benchmark numbers, training dataset details, or sample outputs at launch, leaving practitioners to test inference behavior and compare against the v4 baseline themselves.

The Shahrazad series has carved out a niche among users seeking open-weight models without safety tuning, and the v4.1 increment keeps that momentum alive. What remains unclear is what specific improvements v4.1 brings over v4 — whether the fine-tune targets coherence, instruction-following, or domain-specific knowledge. Future releases should clarify training deltas and ship with eval numbers to help users decide whether the upgrade is worth adopting.