tandrey/Ornith-1.5-9B
Ornith-1.5-9B is a 9 billion parameter dense language model developed by the Ornith Team, extending the Ornith-1.0 architecture (based on Qwen3.5 and Gemma4). This model is distinguished by its end-to-end self-improvement loop, continuously generating new tasks, discovering solution strategies, and refining its policy through reinforcement learning. Optimized for reasoning and agentic tasks, it demonstrates strong performance in coding benchmarks and supports a 262,144 token context window, making it suitable for efficient single-GPU deployment and edge devices.
Loading preview...
Ornith-1.5-9B Overview
Ornith-1.5-9B is a 9 billion parameter dense model from the Ornith-1.5 family, developed by the Ornith Team. It represents a significant advancement in self-improving foundation models, moving beyond fixed human-curated tasks to continuously generate new training tasks, discover effective solution strategies, and optimize its policy via reinforcement learning. This model builds upon the Ornith-1.0 architecture, which was initially developed on Qwen3.5 and Gemma4 with extensive pre-training and fine-tuning.
Key Capabilities
- Advanced Self-Improvement: Features an end-to-end self-improvement loop for task generation, scaffold construction, and solution rollouts.
- Strong Coding Performance: Achieves competitive results across various coding benchmarks, including Terminal-Bench 2.1 (46.2 Terminus-2, 47 Claude Code), SWE-bench Verified (70.6), SWE-bench Pro (47.5), and SWE-bench Multilingual (54.4).
- Enhanced Reasoning: Demonstrates improved reasoning capabilities, scoring 20.2 on HLE (no tools) and 30.5 (with tools), and 86.4 on GPQA Diamond.
- Agentic Functionality: Excels in agentic tasks, with scores like 54.2 on MCP-Atlas, 41.2 on Toolathlon-Verified, and 66.5 on ClawEval. It supports OpenAI-compatible tool calling.
- Long Context Window: Handles a substantial context window of up to 262,144 tokens, extendable to approximately 1 million tokens using YaRN scaling.
- Efficient Deployment: Designed for efficient deployment on a single 80GB GPU and has a quantized Mobile variant for edge devices.
Good for
- Coding and Software Engineering: Ideal for terminal-based coding agents, understanding large codebases, and automating development tasks.
- Complex Reasoning Tasks: Suitable for applications requiring advanced logical deduction and problem-solving.
- Agentic Workflows: Excellent for building AI agents that utilize tools and interact with environments.
- Resource-Efficient Deployment: A strong choice for scenarios requiring powerful language models on single GPUs or mobile devices.