✦
Summary
K2-Type-0.9B is a compact decision model that returns typed probabilities rather than generating text. One forward pass evaluates a shared state against multiple independent questions, using a block-causal mask so questions do not influence each other.
Key Features
Text or JSON state input.
Boolean probabilities.
Multiple-choice decisions.
Ordered scoring levels.
Multiple questions per pass.
Question-independent attention.
A local CUDA server.
Apache-2.0 model release.
About K2-Type-0.9B
Developers can build calibrated classifiers, choice selectors, and scoring components without paying the latency of an autoregressive answer. Core capabilities include Text or JSON state input, Boolean probabilities, Multiple-choice decisions, and Ordered scoring levels.
Public code or model weights are linked from the project, with setup instructions for running or adapting the system. Additional characteristics include Multiple questions per pass, Question-independent attention, A local CUDA server, and Apache-2.0 model release.