Simon Willison · 2026-06-29 · notable
Simon Willison: Ornith-1.0 — hands-on with the open-weights coding model
Simon Willison runs DeepReinforce's new Ornith-1.0 — an MIT-licensed coding model family (9B, 31B, 35B MoE, 397B MoE) built on Gemma 4 and Qwen 3.5 — through LM Studio and a Pi agent loop on a Datasette codebase.
Simon Willison's hands-on first look at DeepReinforce's MIT-licensed Ornith-1.0 — pelican test, agent loop, and the variant lineup.
What is it?
Simon Willison's June 29, 2026 post is a hands-on first look at Ornith-1.0, the first model release from DeepReinforce. The open-weights family ships as 9B Dense, 31B Dense, 35B MoE, and 397B MoE variants, all MIT-licensed and built on Gemma 4 and Qwen 3.5 pretrained checkpoints.
How does it work?
Willison downloaded the ornith-1.0-35b-Q4_K_M.gguf (20GB) quantization and loaded it into LM Studio, then wired it up to Pi for tool-use. He pointed the resulting agent at a Datasette repository to test code search and navigation, and asked the model to draw the SVG pelican he uses as a recurring vibes benchmark — that prompt returned at 103 tokens per second.
Why does it matter?
Open-weights coding models keep arriving faster than most people can keep up with, so a quick hands-on from a trusted reviewer is useful triage. Willison confirms the model runs in LM Studio off the GGUF, handles real agentic tool use on a Python codebase, and characterizes the pelican output as 'a little bit mangled but the pelican is clearly a pelican' — enough to put Ornith on the shortlist of open coding models worth pulling locally.
Who is it for?
Open-source agent builders deciding whether to pull Ornith locally
Try it
https://simonwillison.net/2026/Jun/29/ornith/