Models I've trained and released, and the research questions behind them. Most of it is about memory: what a model should keep, when it should look something up, and how small it can be while still being useful on a laptop.
An open-weight model built to be an on-device Mac agent.
The release reports 97.7% action match on held-out MacBench in BF16. That measures individual actions, not success across a complete Mac workflow.
Receipts and invoices in, structured tool calls out.
Field match improved from 0.12 to 0.99 on OCR text from 100 held-out CORD receipts, using the release prompt and tool schema. Image-input accuracy was not tested in that evaluation.
A small language model that decides when to retrieve.
The release reports 82.5% reflection-token accuracy, a measure of retrieval-control tokens rather than overall answer correctness.
Co-authored with Naga Sri Arvapalli.
Pre-Tokenizer Latent Memory Routing.
A controlled study with 9.3M parameters reported 7–10% lower loss in evidence regions. Results at larger model sizes remain an open question.
A personal-assistant harness for macOS.
The harness provides 24 macOS connectors behind a permission policy.
All released weights live under huggingface.co/nuroai. Numbers are quoted as they appear in each model card, eval log or paper.