Nace.AI has open-sourced Drex-Max v1, a dense 27B multimodal decision model released under the Open RAIL-M license with a 262K context window. Unlike standard LLMs, it does not generate tokens; instead, it accepts context states including text, JSON, and high-resolution images alongside named questions, returning probabilities over candidate answers via a 255-way readout head in a single forward pass. The model supports local inference on Apple Silicon and NVIDIA GPUs via Python.