source: simon willison: ornith-1.0: self-scaffolding llms for agentic coding
level: technical
deepreinforce has released ornith-1.0, a new open weights model family under the mit license. the models are built on top of pretrained gemma 4 and qwen 3.5 architectures, with sizes including 9b dense, 31b dense, 35b mixture of experts, and 397b mixture of experts. they achieve state-of-the-art performance among open-source models of comparable size on coding benchmarks. both underlying models use apache 2.0 licensing, making the combination legally compatible.
the model is designed for agentic coding, meaning it can use tools and make multiple function calls to complete tasks. early tests using the 35b version quantized to 4-bit show it can handle complex multi-step requests, such as finding specific code in a codebase and then performing follow-up actions. it also generates images, like a pelican drawing, at high speed. the model runs locally via lm studio and can be connected to coding assistants.
little is known about deepreinforce, the organization behind the model. their earliest public work appears to be a june 2025 paper on cuda optimization using contrastive reinforcement learning. the release of ornith-1.0 marks their first model launch, contributing a capable open-source option for developers who want local, agentic coding assistants without relying on proprietary apis.
why it matters: it provides a high-performance, open-source alternative for local ai coding agents, reducing dependency on cloud services and enabling private, customizable development workflows.
source: simon willison: ornith-1.0: self-scaffolding llms for agentic coding