Research

Milestone

Aug 21, 2026

Project Kostant: Fast, Fully Local LLMs on Phone Processors

What would it take to build an Ollama/vLLM for phones? Kostant's central result: no single compression trick is sufficient — the winning stack is co-designed across models, kernels, and scheduling.

Release

Aug 15, 2026

Meridian-AC-Nano: a 751.6M-parameter academic assistant

The production cut of the Meridian recipe — multi-task academic writing, proofreading, and LaTeX-aware processing in a small model that runs locally.

Publication

Aug 9, 2026

Meridian 0.1: Why Data Design Beats Method Innovation

A 135M-parameter case study in LaTeX-aware academic proofreading. We show curated data and task decomposition delivering a 2.2× ROUGE-L gain — the same recipe scaling to Meridian-AC-Nano at 751.6M parameters.

Release

Jun 16, 2026

Cobentt: deep research, documents, and smart contracts in one workspace

One AI workspace for documents, slides, sheets, deep research, chat, websites, and smart contracts — now live, with a mobile playground in testing.

Conclusion

Jun 1, 2026

RRbench (ResearchBench): Trust-first verification for scientific research

Citations and journal reputation are insufficient for modern AI-driven science. RRbench combines AI analysis, artifact integrity, expert review, and community validation into one transparent Trust Score.

Milestone

May 25, 2026

Openbentt Beta: the local-first AI research workspace

LaTeX editing, PDF preview, benchmarking, and fine-tuned small models — with inference, history, and attachments staying on your machine.