Ornith 1.0

Model Overview

Ornith 1.0 is an open-weights family of self-scaffolding models optimized for agentic coding, released by DeepReinforce in June 2026. Post-trained on Gemma 4 and Qwen 3.5 base weights, the Ornith family utilizes a reasoning-first approach where models begin responses with a <think> block to strategize before generating code or tool calls. The model family spans compact 9B and 31B dense models, up to massive 35B and 397B Mixture-of-Experts (MoE) frontier sizes, with context windows expandable up to 1M tokens via YaRN RoPE scaling.

Capabilities

Unlike conventional coding agents that rely on fixed, human-designed harnesses to manage execution loops and tool calls, Ornith 1.0 features a self-scaffolding capability trained via reinforcement learning (RL). This allows the model to learn and refine its own internal processes, dynamically adapting to complex software engineering tasks. It also incorporates a three-layer defense system (fixed trust boundary, deterministic monitor, and frozen LLM judge) to prevent reward hacking during RL training.

Example Use Cases

• Autonomous Software Engineering: Functioning as a local agentic coder capable of iterative planning, code generation, and self-correction.
• Local Private Development: The smaller 9B/31B models and GGUF quantization formats allow developers to run the agent locally for privacy-preserving code generation.
• Agentic Workflows: Seamless integration into agent frameworks like vLLM, SGLang, and Transformers using OpenAI-style endpoints.

Performance & Benchmarks

The flagship 397B MoE model has demonstrated frontier-level performance, reportedly outperforming leading models like Claude Opus 4.7 on rigorous coding and agentic benchmarks, including SWE-Bench Verified and Terminal-Bench 2.1.

Intended Use & Limitations

Ornith 1.0 is intended for developers, researchers, and enterprises needing a powerful, open-weights coding agent that can operate autonomously. Because it dynamically generates its own scaffold, it can sometimes pursue unpredictable problem-solving strategies. Despite its three-layer defense system against reward hacking, users should exercise standard precautions and review generated code before executing it in production environments.

About DeepReinforce

DeepReinforce is an AI research institution focused on advancing reinforcement learning techniques for software engineering and reasoning models.