The 11-layer mmBERT-small derivative ships as a 183 MB ONNX file under MIT; retain its copyright and permission notices when redistributing. Seven fixed outputs also flag tools, vision, long context and follow-ups. It reads text only, up to 512 tokens with 600 characters of conversation context, and ignores question wording. It is not Jev or a general-purpose decision model. Training requests and most test labels are synthetic; Decision Models' reported advantage over Jev 1.13 is task-specific, with CPU inference timed separately from hosted API round trips.
How you can use it
Start with the model card's local Python code example if you want your app to choose an AI model for each request. A developer can connect its labels to your list of models, matching the task to a model with the right abilities.
Keep that selection rule in your app; this model does not answer the request or send it onward. Test with prompts you know before trusting its choices. It only reads a short piece of text, cannot inspect images, and was trained on requests written by other AI models.
