AItonomy/PhAI-IDE-72B

TEXT GENERATIONPricing:Input $12 / Output $20Concurrent Unit Cost:4Model Size:72.7BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Sep 15, 2026License:otherArchitecture:Transformer0.0K Featherless Exclusive Cold

AItonomy/PhAI-IDE-72B is a 72.7 billion parameter model from the PhAI-IDE family, developed by AItonomy. This model is supervised fine-tuned using ms-swift on Codex trajectories, specializing in scientific coding and tool interaction. It demonstrates strong performance in scientific code repair and outperforms several reference models on benchmarks like AQuA-RAT and ARC-Challenge, making it suitable for complex scientific programming tasks.

Loading preview...

PhAI-IDE-72B: Specialized for Scientific Coding and Tool Interaction

PhAI-IDE-72B is the largest model in the PhAI-IDE family, developed by AItonomy, and is specifically designed for advanced scientific coding and seamless interaction with various tools. This 72.7 billion parameter model is built upon the Qwen2.5-72B-Instruct base and has undergone supervised fine-tuning using ms-swift.

Key Capabilities

  • Scientific Code Repair: Excels at localized scientific-code repair on familiar codebases, as demonstrated by significant gains over base models in the ScienceAccelBench performance (e.g., +12.50 pp in MITgcm-biogeo for the 9B variant).
  • Tool Interaction: Trained on Codex trajectories that capture code inspection, tool use, and responses to execution feedback, enabling robust interaction with external tools.
  • Strong Benchmark Performance: Outperforms several reference models of similar size on critical benchmarks:
    • AQuA-RAT: Achieves 77.56%, a +46.24 pp difference over Llama-2-70B-Chat.
    • ARC-Easy: Scores 84.64%, +8.14 pp over Llama-2-70B.
    • ARC-Challenge: Reaches 64.42%, +4.92 pp over Llama-2-70B.

Good For

  • Scientific Computing: Ideal for applications requiring code generation, debugging, and repair within scientific domains.
  • Automated Development Environments: Suitable for integration into IDEs or platforms that benefit from AI assistance in coding and tool orchestration.
  • Research and Development: Valuable for researchers and developers working on complex scientific problems that involve extensive coding and numerical verification.

Finetunes
1 models