Ornith-1.0: Self-Scaffolding LLMs for Agentic Coding

Ornith-1.0: Self-Scaffolding LLMs for Agentic Coding. That is an attention-grabbing new open weights (MIT licensed) mannequin, the primary mannequin launch from DeepReinforce.

[…] with variants together with 9B Dense, 31B Dense, 35B MoE, and 397B MoE. Constructed on high of pretrained Gemma 4 and Qwen 3.5, it achieves state-of-the-art efficiency amongst open-source fashions of comparable measurement on coding benchmarks.

So far as I can inform the licenses of these underlying fashions is suitable with getting used on this method – Gemma 4 is Apache 2.0 licensed (and never certain by the janky further Gemma Phrases of Use that troubled the earlier Gemma fashions) and Qwen 3.5 is Apache 2.0 licensed as properly.

I have been working the mannequin utilizing LM Studio and the ornith-1.0-35b-Q4_K_M.gguf (20GB) GGUF, hooked as much as Pi. Preliminary impressions are excellent – it appears to have the ability to run the agent harness over many device calls in a proficient method.

Here is a terminal session the place I requested it to “discover the code that decodes the actor cookie” after which “discover the code that opens the insert dialog when thebutton is clicked” in opposition to a Datasette checkout, which it dealt with with ease.

I additionally had it draw this pelican, which got here out at 103 tokens/second:

It is slightly bit mangled however the pelican is clearly a pelican.

I could not discover a lot details about DeepReinforce themselves. The earliest paper I might discover from the was CUDA-L1: Enhancing CUDA Optimization by way of Contrastive Reinforcement Studying from June 2025.