Summary
Bart Slodyczka tests Ornith 1.0, a newly released open-source agentic coding model family from Deep Reinforce AI (San Francisco), focusing on the 9-billion parameter dense variant running locally on a 16GB M4 Mac Mini. Ornith 1.0 is built on top of Gemma 4 and Qwen 3.5 base models and ships in four sizes — 9B dense, 31B dense, 35B MoE, and 397B MoE — with the 9B, 35B, and 397B variants available as open weights today via LM Studio.
The review covers model setup in LM Studio with Q4 quantization (~5.63 GB), integration with PI Agent as a coding harness (chosen over Claude Code for faster response latency on constrained hardware), and real-world coding task performance — specifically building a tower defense game in a single HTML file, the same benchmark previously run against Qwen 3.6 35B. Measured inference speed runs 16–18 tokens per second on the M4 Mac Mini with 12GB allocated to the model. The video also references Ornith’s published benchmark results on TerminalBench 2.1, SWE-bench Verified, and the OpenClaw evaluation, with honest skepticism about how well those translate to from-scratch generation tasks.
The verdict is nuanced: the 9B model shows capable reasoning and structured thinking even for a small parameter count, but its ceiling for complex agentic builds is noticeably lower than the 35B MoE variant. The video is an honest guide for developers evaluating whether Ornith 1.0 is worth running locally on consumer hardware, complete with configuration specifics and side-by-side output comparisons.
📺 Source: Bart Slodyczka · Published June 29, 2026
🏷️ Format: Review
