Drex Max offers local answers to structured questions

The model card describes a different task from writing a chatbot reply: Drex Max returns probabilities for the supplied options. Its Python implementation accepts text and images on NVIDIA hardware. The documented Apple silicon paths run the text backbone. The affine Q4 version requires at least 24 GB of memory; Q8_0 requires 48 GB. Those requirements distinguish the local versions before installation.

The request specification shows how a developer defines the job. A state field holds the document as text, an object or a list, while named questions specify what to assess. Its support-ticket example asks which team should handle a message, whether a refund was explicitly requested and how urgent the ticket is. These produce a choice, a yes-probability and a weighted rating. The confidence field has different definitions for choices and ratings; the local rating formula approximates the hosted service’s.

The separate hosted-service guide requires an explicit model choice to reach Drex Max: omitting it selects Drex v2.0. The original Max and Max v1.1 are separate pinned versions. Images belong in their own array, seen by every question, rather than inside the state or questions. A URL supplied as text does not cause an image download. These are hosted request rules, not local installation instructions.

MarkTechPost’s November 1 coverage links the weights and local-running documentation and discloses sponsorship by Nace AI.