Badger Code: Building a Local Coding Agent on a Single GPU

Badger Code is a challenge in the 2026 Machine Learning Marathon (MLM26). Teams build the best open coding agent they can on a single GPU – no proprietary models, no giant clusters – scored on Terminal-Bench 2.1. Terminal-Bench measures whether an agent can use shell commands, inspect files, read error messages, and iterate until the task is complete, connecting models to a sandboxed terminal for reproducible evaluation.

Challenge design

• Task: Assemble an agent (open-weights model + harness) that solves real terminal-based tasks end to end.
• Constraints: Open models only, running on a single GPU – efficiency and clever harness design matter as much as raw model size.
• Evaluation: Terminal-Bench 2.1 task completion rate in a sandboxed, reproducible execution harness.

Links

• Kaggle challenge: Badger Code

Related resources

• Benchmark: SWE-bench: Evaluating AI on Real-World Software Engineering: The most widely used benchmark for AI coding agents – useful background on how agentic coding is evaluated.
• Notebook: Understanding Quantization and Precision: Learn how quantization (e.g., 4-bit) reduces model size and memory requirements – essential for fitting capable open models on a single GPU.
• Compute: National Research Platform (NRP): Free open-weight LLM endpoints (with a Python code-along) plus free GPU compute for UW-Madison researchers – prototype your agent against the hosted API, then develop and benchmark on your own GPU pod.

Comments