DecisionTune 1.0 Released as a Fast 395M Encoder for Local AI Decision-Making
DecisionTune 1.0 has been released under the Apache-2.0 license, introducing a 395M parameter encoder model based on ModernBERT-large paired with a 4 KB scoring head for local decision-making. Rather than generating text, it processes state and option inputs in a single pass to output option probabilities in roughly 10 ms per short query on Apple Silicon via MLX. In agentic AI systems, running full autoregressive LLMs for routine tasks like tool routing or ticket classification introduces unnecessary latency and compute costs. DecisionTune offloads these frequent, small choice tasks to a lightweight local encoder, enabling near-instantaneous decision-making without external API dependencies. The model features an 8,192-token context limit that strictly rejects oversized inputs, supporting PyTorch, ONNX, and Apple MLX backends with a 99.85% cross-backend answer parity. However, it requires well-described option strings to perform accurately, is limited to English, and does not perform general knowledge reasoning or mathematical calculations.
## BACKGROUND
ModernBERT is an updated encoder-only transformer model designed for improved efficiency, speed, and longer context window processing compared to classical BERT architectures. Apple MLX is an open-source machine learning framework optimized for Apple silicon's unified memory architecture, allowing high-performance local inference. Unlike generative autoregressive models that predict text one token at a time, encoder models evaluate context sequences all at once, making them significantly faster for classification and scoring tasks.