Willison Runs an Open Model That Writes Its Own Training

Simon Willison ran Ornith-1.0, an open coding model that writes its own reinforcement-learning training scaffold, on a single machine, per his June 29 write-up.

That's the case DeepReinforce is making to developers wary of sending code to a hosted API. The full agent loop, tool calls included, runs from a 35B mixture-of-experts model packed into a roughly 20GB file, on one high-memory machine you own.

Willison loaded the GGUF file in LM Studio and wired it into his own Pi coding harness. In his tests, Ornith-1.0 drove an agent loop across many tool calls without stumbling, he wrote. The model builds on two pretrained bases, Gemma 4 and Qwen 3.5, and ships under an MIT license.

The self-scaffolding part is the novelty: the model writes the training harness that guides its own reinforcement-learning improvement. It's a credible local fallback for agentic coding, especially for anyone who read the steganography story and wants an alternative to hosted models.