ornith-ai/Ornith-1.5-9B
Ornith-ai's Ornith-1.5-9B is a 9 billion parameter dense language model, building upon Qwen3.5 and Gemma4 architectures. It features an advanced self-improvement loop that continuously generates new training tasks and optimizes solution strategies through reinforcement learning. This model excels in coding, reasoning, and agentic tasks, demonstrating strong performance across various benchmarks, and is designed for efficient single-GPU deployment with a 32,768 token context window.
Loading preview...
Ornith-1.5-9B: Self-Improving Language Model
Ornith-1.5-9B, developed by ornith-ai, is a 9 billion parameter dense model that represents a significant advancement in self-improving foundation models. It extends the previous Ornith-1.0 by enhancing its self-improvement loop to jointly optimize task generation, scaffold construction, and solution rollouts, moving beyond reliance on fixed human-curated tasks. This model is built on top of Qwen3.5 and Gemma4, incorporating continued pretraining, mid-training, and post-training.
Key Capabilities
- Advanced Self-Improvement: Continuously generates new training tasks and discovers effective strategies through reinforcement learning.
- Strong Coding Performance: Achieves high scores on benchmarks like SWE-bench Verified (70.6), SWE-bench Pro (47.5), and Terminal-Bench 2.1 (46.2 Terminus-2, 47 Claude Code).
- Enhanced Reasoning: Demonstrates robust reasoning capabilities, scoring 30.5 on HLE (with tools) and 86.4 on GPQA Diamond.
- Agentic Task Proficiency: Excels in agentic benchmarks such as MCP-Atlas (54.2), Toolathlon-Verified (41.2), and ClawEval (66.5).
- Efficient Deployment: Designed for efficient single-GPU deployment and has a quantized mobile variant (Ornith-1.5-9B-Mobile).
- Long Context Window: Supports a native context window of 262,144 tokens, extendable to approximately 1 million tokens using YaRN scaling.
- Tool-Use and Reasoning Parsing: Features built-in reasoning and tool-call parsers, surfacing chain-of-thought in a
reasoning_contentfield and tool calls in OpenAI-styletool_calls.
Good for
- Coding and Software Engineering: Ideal for tasks requiring code generation, understanding, and debugging, as evidenced by strong SWE-bench and Terminal-Bench results.
- Complex Reasoning: Suitable for applications demanding advanced logical deduction and problem-solving.
- Agentic Workflows: Excellent for integrating into agent frameworks due to its robust tool-calling capabilities and performance on agentic benchmarks.
- Resource-Efficient Deployment: A good choice for scenarios requiring powerful language models on single GPUs or edge devices.
- Long-Context Applications: Beneficial for processing and generating content within very large context windows, up to 1 million tokens with YaRN scaling.