Ornith-1.0: Self-Scaffolding LLMs for Agentic Coding

2026-06-29 · Source: Simon Willison's Weblog · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Software Development & Engineering · Depth: Intermediate, quick

Summary

DeepReinforce has released Ornith-1.0, a new open-weights, MIT-licensed large language model designed for agentic coding. Available in variants including 9B Dense, 31B Dense, 35B MoE, and 397B MoE, it is built upon Apache 2.0 licensed Gemma 4 and Qwen 3.5. Ornith-1.0 reportedly achieves state-of-the-art performance among open-source models of comparable size on coding benchmarks. Initial testing with the 20GB "ornith-1.0-35b-Q4_K_M.gguf" variant via LM Studio and Pi demonstrated its proficiency in running agent harnesses over multiple tool calls. Examples include successfully locating specific code within a Datasette checkout and generating an image at 103 tokens/second. DeepReinforce, the developer, appears to be a new entity, with its earliest identified paper, "CUDA-L1: Improving CUDA Optimization via Contrastive Reinforcement Learning," dated June 2025.

Key takeaway

For AI Engineers evaluating open-source models for agentic coding workflows, Ornith-1.0 presents a compelling option. Its reported state-of-the-art performance on coding benchmarks, coupled with its MIT license and Apache 2.0 licensed foundations, makes it suitable for commercial deployment. You should consider testing the 35B MoE variant for tasks requiring multi-step tool interactions, such as automated code discovery or complex generation, to assess its fit within your existing infrastructure and project requirements.

Key insights

Ornith-1.0 demonstrates strong agentic coding capabilities using self-scaffolding LLMs built on open-source foundations.

Principles

• Open-source LLMs can achieve SOTA coding performance.
• Compatible licensing enables model stacking.
• Agentic LLMs excel at multi-step coding tasks.

Method

The model runs an agent harness over many tool calls, enabling proficient execution of complex coding tasks like code discovery and generation.

In practice

• Use "ornith-1.0-35b-Q4_K_M.gguf" for local agentic coding.
• Integrate with LM Studio and Pi for testing.
• Apply for code search and image generation tasks.

Topics

• Agentic AI
• Large Language Models
• Code Generation
• Open-source Models
• Gemma 4
• Qwen 3.5
• DeepReinforce

Best for: Research Scientist, Machine Learning Engineer, AI Engineer, AI Scientist

Related on AIssential

• This Week in Open Models: Tiny LFM2.5, Ornith-1.0, and GLM-5.2 REAP — Small models can achieve disproportionate instruction-following performance, while advanced training methods enable self-improving agentic coding and efficient MoE pruning.
• GLM 5.1 Agentic Coding Test with OpenCode | New Best Open Coding LLM? | Live Test — GM 5.1 excels in agentic coding and project understanding, outperforming some established models in specific benchmarks.
• : Structuring Your Natural Language SOPs into Tailored Ambiguity-Resolved Code Templates — A multi-agent LLM framework effectively transforms ambiguous SOPs into structured, executable code with high accuracy.
• North Mini Code - Cohere — A compact 3B active parameter MoE model excels in agentic coding and local deployment.
• We reimplemented Claude Code entirely in Python — open source, works with local models — Claw Code Agent offers a Python-native, open-source alternative to Claude Code for local agentic development.
• Introducing Command A+: Making sovereign agentic capabilities available to all - Cohere — Cohere's Command A+ democratizes enterprise-grade agentic AI through an efficient, open-source MoE model.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Simon Willison's Weblog.