manojpanda/peta1

VISIONPricing:Input $0.4 / Cached $0.08 / Output $0.8Concurrent Unit Cost:1Model Size:7.9BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Oct 14, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Cold

Peta1 is a 7.9 billion parameter autonomous software engineering agent checkpoint developed by manojpanda, based on Google's Gemma 4 E4B-it model. It was post-trained using a three-stage curriculum involving Tree-of-Reasoning (ToR) SFT, Capability Frontier Discovery, and Verifiable Outcome GRPO. This specialized training significantly enhances its ability to solve software engineering tasks, achieving a 33.67% resolution rate on SWE-bench Lite, a 25.33 percentage point gain over the base model, and excels at autonomous tool navigation.

Loading preview...

Peta1: Autonomous Software Engineering Agent

Peta1 is a specialized autonomous software engineering agent developed by manojpanda, built upon the google/gemma-4-E4B-it base model (4.3B active parameters). It has undergone a unique three-stage post-training curriculum designed to significantly improve its problem-solving capabilities in software engineering tasks.

Key Capabilities

  • Enhanced Reasoning: Utilizes Tree-of-Reasoning (ToR) Supervised Fine-Tuning (SFT) to enforce hypothesis formulation, environment simulation, and branch pruning, leveraging Gemma 4's native latent thinking tokens.
  • Targeted Optimization: Employs Capability Frontier Discovery to focus training on informative tasks, avoiding trivial or impossible ones that can lead to policy gradient collapse.
  • Verifiable Outcomes: Optimized with Group Relative Advantage Policy Optimization (GRPO) using verifiable rewards, ensuring robust and effective learning.
  • Superior Bug Resolution: Achieves a 33.67% resolution rate on the full official SWE-bench Lite benchmark (101/300 tasks), representing a +25.33 percentage point gain over the base Gemma 4 E4B model.
  • Improved Tool Navigation: Demonstrates a significant improvement in autonomous tool navigation, achieving 25.0% success in turn 1 of agentic evaluations, compared to 0.0% for the base model.

Good for

  • Automated software bug fixing and resolution.
  • Tasks requiring autonomous tool navigation and environment interaction.
  • Developers seeking a compact model with strong latent code comprehension for agentic workflows.
  • Research into advanced reasoning and verifiable outcome optimization for LLMs in software engineering.

Finetunes
1 models