nuroai/eliot-9b
Eliot 9B by Nuro AI Labs is a 9 billion parameter, Qwen3.5-9B fine-tuned, open-weight tool-calling model specifically designed for Apple-platform agentic control. It operates a Mac through a guarded macOS harness by interpreting Accessibility trees and emitting structured actions. This model excels at local-first assistants, Mac automation, and agentic desktop control, focusing on observable and auditable actions.
Loading preview...
Eliot 9B: An Apple-Platform Agent Model
Eliot 9B, developed by Nuro AI Labs, is a 9 billion parameter model fine-tuned from Qwen/Qwen3.5-9B with a 32768 token context length. Unlike most LLMs trained for chat, Eliot is engineered to act as an agent on macOS. It interprets macOS Accessibility trees, reasoning over UI elements to emit structured actions like click, type, open_app, read, and web_search.
Key Capabilities & Differentiators
- macOS Accessibility Control: Operates via structured UI trees, not raw pixels, enabling faster prompting and easier inspection.
- Predictable Agent Loop: Designed for one tool call per turn, simplifying integration into harnesses and product workflows.
- Apple Silicon Optimization: A dedicated MLX 4-bit build (
nuroai/eliot-9b-mlx-4bit) is available for efficient local execution on Macs. - Harness-First Safety: Built with the expectation that a runtime harness will intercept and guard destructive or sensitive actions, ensuring auditable and controllable operations.
- Local-First Design: Ideal for assistants that process user data locally, enhancing privacy and reducing reliance on remote services.
Intended Use Cases
Eliot 9B is particularly well-suited for:
- macOS assistant research and prototyping.
- Developing local-first desktop automation solutions.
- Tool-calling and computer-use harness development.
- Experiments with Apple Silicon assistants.
- Evaluating Accessibility-tree-based agent interfaces.