Ornith-1.0: Self-Scaffolding LLMs for Agentic Coding

Why it matters: An unexpected open-source agent coding model Simon Willison, the prolific open-source tool maker, recently shared his first impressions of Ornith-1.0 on his blog. Released by the new team DeepReinforce, this model targets the cutting-edge field of agentic coding. Rather than training from scratch, it uses a self-scaffolding approach, fine-tuning on top of Gemma 4 and Qwen 3.5. This raises an intriguing question: as open-source base models become more powerful, can we build specialized agent models through clever combinations and secondary training?

What makes Ornith-1.0 special? Ornith-1.0 is not just another coding model. Its core design goal is self-scaffolding — enabling the model to self-correct and advance tasks during code generation, tool calling, and multi-step reasoning. This is evident in its strong agent harness capability: Simon ran the 35B version locally via LM Studio and asked it to "find the code that decodes the actor cookie" and then "find the code that opens the insert dialog when the button is clicked" in a Datasette codebase. The model not only completed the tasks accurately but also made multiple tool calls smoothly. Even more striking, it generated a complex SVG drawing of a pelican — slightly mangled but clearly a pelican — at 103 tokens per second.

Technically, the model cleverly avoids licensing pitfalls. Both Gemma 4 and Qwen 3.5 use the Apache 2.0 license, making their combination for fine-tuning legally sound. This serves as a reminder to the open-source community: future model innovation need not always start from scratch. Picking a permissively licensed base for secondary development may be a faster, more compliant path.

Trend: Agent models are becoming specialized Ornith-1.0 reveals a deeper trend: after the race for general-purpose LLMs, specialized agent models for particular workflows are emerging. These models stop trying to know everything and instead focus on interacting with tools, understanding codebases, and executing multi-step operations. Like the models behind commercial products such as Devin, Ornith-1.0 shows that the open-source community can keep pace, even attracting developers sensitive to data privacy and cost with more open licenses and the advantages of local deployment.

How you can use it If you're a developer, Ornith-1.0 offers several immediate possibilities: first, run an agent coding assistant locally, avoiding API latency and cost; second, pair it with terminal tools like Pi to let AI help you explore large codebases right from the command line; third, use it as a base model for your own agent applications through further fine-tuning. The GGUF quantized version shared by Simon requires only 20 GB of VRAM, making it runnable on consumer GPUs. While the 35B model may not match GPT-4 in overall capability, it is already useful for specific tasks.

Counterintuitive insight: Licensing is not a barrier but a new ecosystem enabler Many assume that open-source model licenses are a tangled mess that makes combinations risky. But Ornith-1.0 proves that by paying attention to license compatibility, permissive licenses like Apache 2.0 fully support secondary innovation. This opens the door to model mashups — you can combine model weights just as you combine software libraries, rapidly building best-of-breed vertical solutions.