AItonomy/PhAI-IDE-4B

VISIONPricing:Input $0.4 / Cached $0.08 / Output $0.8Concurrent Unit Cost:1Model Size:4.5BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Sep 15, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

AItonomy/PhAI-IDE-4B is a 4.5 billion parameter model from the PhAI-IDE family, fine-tuned for scientific coding and tool interaction. Based on Qwen3.5-4B, it leverages Codex trajectories for training and excels at scientific code repair and numerical verification tasks. This model is optimized for developers working on scientific computing and environments requiring precise code generation and interaction.

Loading preview...

PhAI-IDE-4B: Scientific Coding and Tool Interaction Model

PhAI-IDE-4B is a 4.5 billion parameter model developed by AItonomy, part of the PhAI-IDE family which also includes 9B and 72B variants. It is built upon the Qwen3.5-4B base model and has been specifically fine-tuned using ms-swift with a dataset of Codex trajectories sourced from ScienceIDE. This specialized training focuses on code inspection, tool use, and responding to execution feedback, making it highly effective for scientific coding tasks.

Key Capabilities

  • Scientific Code Repair: Demonstrates significant gains in localized scientific-code repair, outperforming its base model (Qwen3.5-4B) by +33.33 percentage points on tasks like PLUTO-Particles-Dust.
  • Tool Interaction: Trained to capture exec / wait interaction formats, enabling effective tool use within coding environments.
  • Numerical Verification: Utilizes a numerical-equivalence verifier during its training process, enhancing its ability to handle tasks requiring precise numerical outcomes.
  • Strong Benchmark Performance: Achieves competitive scores on various benchmarks, including 97.60% on BBH multistep-arithmetic-two, showing a +44.40 difference compared to Llama-3.2-3B-Instruct.

Good For

  • Scientific Computing: Ideal for applications involving scientific code generation, analysis, and debugging.
  • Automated Code Repair: Useful for automating the repair of scientific codebases, especially in environments like PLUTO-Particles-Dust.
  • Interactive Development Environments: Designed for integration into IDEs or systems that require models to interact with tools and respond to execution feedback.
  • Research and Development: Provides a robust foundation for further research into AI-assisted scientific programming and tool-augmented language models.

Finetunes
1 models