FrontisAI/Frontis-MA1-35B

TEXT GENERATIONConcurrent Unit Cost:3Model Size:35.1BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Jul 30, 2026License:cc-by-nc-4.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Cold

FrontisAI/Frontis-MA1-35B is a 35.1 billion parameter execution-grounded research model, based on Qwen3.6-35B-A3B, developed by FrontisAI. It is specifically post-trained for evolutionary machine learning engineering (MLE) and AutoResearch, utilizing execution-grounded supervised fine-tuning and reinforcement learning. The model excels at program transformation tasks like drafting, improving, debugging, and recombining code within an evolutionary search loop. It demonstrates significant gains in MLE-Bench Lite performance, improving Medal Average from 39.39% to 60.61% over its base model.

Loading preview...

Frontis-MA1-35B: An Execution-Grounded Model for Evolutionary ML Engineering

Frontis-MA1-35B, developed by FrontisAI, is a 35.1 billion parameter model derived from Qwen3.6-35B-A3B. It is the flagship research model for OpenMLE, an execution-grounded system designed for meta-evolution in machine learning engineering (MLE). This model is uniquely post-trained using execution-grounded supervised fine-tuning and reinforcement learning, enabling it to learn and apply four core program-transformation operators: Draft, Improve, Debug, and Crossover.

Key Capabilities & Differentiators

  • Execution-grounded post-training: SFT examples and RL rewards are derived from programs executed in isolated MLE sandboxes, ensuring practical code generation and refinement.
  • Trainable evolutionary operators: The model's ability to Draft, Improve, Debug, and Crossover code provides a shared action space for both training and test-time evolutionary search.
  • Strong performance gains: On MLE-Bench Lite, Frontis-MA1-35B, when integrated with the OpenMLE-Evo harness, significantly improves the Medal Average from 39.39% to 60.61% and Human Rank from 0.5828 to 0.7647 compared to its base model. With OpenMLE-Evo-Max, it reaches 71.21% Medal Average and 0.8126 Human Rank.
  • Transferability: Demonstrates focused transfer capabilities on NatureBench Lite, improving Surpass-SOTA by 10 percentage points and Match-SOTA by 20 points over its base model.

Ideal Use Cases

  • Research in MLE and AutoResearch agents: Specifically designed for studying and advancing automated machine learning engineering.
  • Iterative improvement of ML experiment code: Generating and refining code for machine learning experiments.
  • Execution-grounded program search: Best utilized within evolutionary search harnesses like OpenMLE-Evo or compatible sandboxed environments.

Finetunes
3 models