Bijan Bowen tests Ornith 1.0 in two local model sizes, including a 35B mixture-of-expertsA mixture-of-experts model contains multiple specialized subnetworks and activates a selected subset for each input instead of using every parameter every time. version with a smaller active parameter count. He compares practical memory requirements and generation behavior before moving to matched coding tasksEvaluation is the systematic assessment of a system's behaviour or results against explicit criteria..
The tests cover repository understanding, browser interfaces, game creation and Blender work. Both models show useful instruction following, while the larger model generally produces more coherent structure, stronger detail and fewer implementation failures.
Bowen treats Ornith as evidence that local coding modelsAn AI coding agent is a tool-using AI system that can inspect, modify, and validate software within a repository. are improving quickly, but not as a reason to ignore hardware limitsAI compute infrastructure is the hardware, facilities, networks, power, cooling, storage, and software used to train and run AI models. or output review. Users gain privacyData privacy governs how personal, confidential, or sensitive information is collected, used, shared, retained, and protected in AI systems. and control while still accepting slower runs and more variable execution than top hosted systems.
Watch on YouTube
