Pq234/Pati

VISIONConcurrent Unit Cost:1Model Size:9BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Jul 1, 2026License:mitArchitecture:Transformer Open Weights Featherless Exclusive Cold

Pq234/Pati is the 9 billion parameter Ornith-1.0-9B model developed by DeepReinforce Team, part of the Ornith-1.0 family of self-improving open-source models. It is specifically optimized for agentic coding tasks, achieving state-of-the-art performance among open-source models of comparable size on benchmarks like Terminal-Bench 2.1, SWE-Bench, NL2Repo, and OpenClaw. This model is designed for efficient single-GPU deployment and excels in tool-calling capabilities.

Loading preview...

Ornith-1.0-9B: Agentic Coding Model

Ornith-1.0-9B is a 9 billion parameter model from the Ornith-1.0 family, developed by DeepReinforce Team. This model is distinguished by its self-improving training framework, which uses Reinforcement Learning to optimize both solution rollouts and the scaffolding that drives them, leading to higher-quality solutions and better search trajectories. It is designed for efficient deployment on a single 80GB GPU.

Key Capabilities

  • State-of-the-Art Agentic Coding: Achieves leading performance among open-source models of similar size on coding benchmarks including Terminal-Bench 2.1 (e.g., 43.1 on Terminus-2), SWE-Bench (e.g., 69.4 on Verified), NL2Repo (27.2), and OpenClaw (63.1).
  • Reasoning Model: Incorporates a <think> ... </think> block for chain-of-thought processing, with serving recipes that parse this into a reasoning_content field.
  • Tool-Calling: Emits well-formed function calls, parsed into standard OpenAI-style tool_calls, making it compatible with various agent frameworks and coding CLIs.
  • Efficient Deployment: The 9B dense model is optimized for single-GPU inference.

Good For

  • Agentic Coding Applications: Ideal for developing and deploying coding agents that require advanced reasoning and tool-use capabilities.
  • Terminal-Based Coding: Optimized for integration with terminal-based coding agents and CLIs like Hermes Agent, OpenClaw, and OpenCode.
  • Research and Development: Suitable for researchers exploring self-improving LLM architectures and agentic AI.
  • Local Inference: Can be served comfortably on a single 80GB GPU using runtimes like vLLM, SGLang, or Hugging Face Transformers.

Finetunes
1 models