AItonomy/PhAI-IDE-9B

VISIONPricing:Input $0.431 / Cached $0.0862 / Output $1.12Concurrent Unit Cost:1Model Size:9BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Sep 15, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

PhAI-IDE-9B is a 9 billion parameter model from AItonomy, fine-tuned for scientific coding and tool interaction. Based on Qwen3.5-9B, it leverages Codex trajectories from ScienceIDE for supervised fine-tuning. This model excels at scientific code repair and numerical verification, demonstrating significant performance gains over its base model and other similarly sized LLMs on benchmarks like ScienceAccelBench, BBH, and MATH-500.

Loading preview...

PhAI-IDE-9B: Scientific Coding and Tool Interaction Model

PhAI-IDE-9B is a 9 billion parameter model developed by AItonomy, specifically fine-tuned for scientific coding and interaction with tools. It is part of the PhAI-IDE family, which also includes 4B and 72B parameter variants. This model is built upon the Qwen3.5-9B base model and utilizes supervised fine-tuning with ms-swift on a unique dataset of Codex trajectories sourced from ScienceIDE.

Key Capabilities

  • Scientific Code Repair: Demonstrates strong performance in repairing scientific code, particularly on task-held-out, localized scientific-code repair with numerical verification.
  • Tool Interaction: Designed to interact effectively with tools, capturing code inspection and responses to execution feedback.
  • Enhanced Reasoning: Shows significant improvements over its base model and other LLMs of similar size on complex reasoning benchmarks, including BBH (multistep-arithmetic-two, word-sorting) and MATH-500.
  • Supervised Fine-Tuning: Benefits from supervised fine-tuning using LoRA across trainable linear layers for three epochs, with final LoRA adapters merged into the full BF16 weights.

Good For

  • Scientific Computing: Ideal for applications requiring code generation, repair, and understanding within scientific domains.
  • Automated Code Assistance: Useful for tasks involving code inspection, debugging, and responding to execution feedback in scientific workflows.
  • Research and Development: Provides a robust foundation for developing intelligent agents that can interact with scientific tools and environments.

Finetunes
1 models