Skip to content
Main Site News Console

Ornith-1.0: Self-Supporting Large Language Model for Agentic Coding

· Simon Willison Translated
技巧LLM

Ornith-1.0: A Bootstrapped LLM for Agentic Coding

This is an interesting new open-weights model (MIT licensed), and the first model released by DeepReinforce.

[…] It comes in 9B Dense, 31B Dense, 35B MoE, and 397B MoE variants. Built on top of pretrained Gemma 4 and Qwen 3.5, it achieves state-of-the-art performance among open-source models of comparable size on coding benchmarks.

As far as I can tell, the licenses on those base models are compatible with this kind of use — Gemma 4 is licensed under Apache 2.0 (and is not subject to those extra, annoying Gemma Terms of Use that affected earlier Gemma models), and Qwen 3.5 is also licensed under Apache 2.0.

I’ve been running this model in LM Studio and connecting to ornith-1.0-35b-Q4_K_M.gguf (20GB) GGUF model via Pi. My first impressions are very positive — it seems quite capable of running an agentic framework with a lot of tool calls in a reasonably competent way.

Here’s a terminal session where I asked it to first “find the code that decodes the actor cookie” in a Datasette checkout and then “find the code that opens the insert dialog when the button is clicked” — it handled both easily.

I also asked it to draw this pelican, and it generated at 103 tokens/second:

Image 1: A cartoon-style illustration of a white pelican, albeit somewhat distorted, with a large beak and orange bill, riding a red bicycle through green hills. The scene includes a blue sky, a yellow sun, and three white clouds, with a few small tufts of grass in the foreground.

A little distorted, but clearly still a pelican.

I couldn’t find much information about DeepReinforce as a company. The earliest paper I could find was their June 2025 publication, CUDA-L1: Improving CUDA Optimization with Contrastive Reinforcement Learning.

Published June 29, 2026 at 4:17 PM

Recent articles