CMF: Millisecond Decision-Making Without Token Generation and Open Weights

CMF has been introduced as a local model for rapid decision-making that operates without token generation, serving as an alternative to the Jev system. The primary function of CMF is to classify requests and make decisions with minimal latency. The model can assess its own confidence level: if confidence is low, it delegates the request to an 'oracle' (an external model) and learns from the result for future iterations. The article details the CMF architecture, compares it with Laya and Jev, and explains how to run it via CLI and API. Performance optimization is a key highlight: on RTX PRO 4000 Blackwell hardware, the median latency was approximately 1.1–1.2 ms. Support for Apple Metal and Vulkan has also been implemented, expanding the model's deployment capabilities across various hardware platforms.