DeepReinforce released Ornith-1.0, a family of open-weights models (MIT licensed) that achieve state-of-the-art coding performance among open-source models of comparable size, with variants ranging from 9B to 397B parameters. This release pushes forward open-source agentic coding capabilities, enabling developers to run powerful coding assistants locally without vendor lock-in, and the self-scaffolding technique may improve reliability in multi-step tasks. The models are built on top of pretrained Gemma 4 (Apache 2.0) and Qwen 3.5 (Apache 2.0), ensuring license compatibility, and include both dense and Mixture-of-Experts variants. The author tested a 35B GGUF quantized model and reported strong performance on agentic coding tasks.
Background
Self-scaffolding refers to the ability of an LLM to autonomously generate and follow its own reasoning steps or tool-use plans, improving performance on complex agentic tasks without external prompts. Mixture of Experts (MoE) is an architecture that activates only a subset of parameters per token, allowing larger effective model sizes with lower computational cost.