This release is a community fine-tune that strips the safety filters from the base MoE design, returning unfiltered responses without refusal behaviors. The underlying sparse architecture activates only about 3 billion of its 35 billion parameters per token, which keeps inference economical while preserving the broader model's capacity. The same weights power both the standard variant and a thinking variant that encourages more deliberate step-by-step reasoning for complex tasks. Open weights under an Apache-2.0 license allow local hosting, and Q8 quantization is reported to run near full quality on consumer hardware through llama.cpp.

On the NanoGPT-hosted variant, a thinking mode is available for tougher coding, multimodal analysis, tool use, and multi-turn chat work, with structured outputs supported. Independent community benchmark reporting on the CrucibleMark platform shows the model clearing routine tasks comfortably but landing below the leader on heavier synthesis work; tool execution ranks close to the average field, while CLI and code quality sit modestly above it. The combination of broad context, multimodal inputs, tool calling, and permissive output behavior makes this a practical fit for developers and researchers who want an open, locally runnable model for agentic experiments and unconstrained content generation.