# GTerm - Small Model Big Bench

- **Event:** [Google I/O Hackathon](https://cerebralvalley.ai/e/google-io-hackathon)
- **When:** Sat, May 23 at 9:00 AM – 10:00 PM (PDT)
- **Where:** Shack15, San Francisco, CA
- **Team:** [Aditya  Advani](https://cerebralvalley.ai/u/ninjaa), [Dominic Domoah](https://cerebralvalley.ai/u/dominicMolt)
- **GitHub:** https://github.com/Phantastic-AI/gterm-bench-hack
- **Demo video:** https://youtu.be/gOtqXY01FS8
- **Gallery:** https://cerebralvalley.ai/e/google-io-hackathon/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/google-io-hackathon/hackathon/gallery/23

How high can Gemini 3.5 Flash climb on Terminal-Bench? GTerm is an attempt to answer this question. We believe we can beat SOTA! Early results are promising.

How it works: GTerm is running a Meta-Harness-inspired self-improving loop on Flash. The original example in the Meta-Harness paper reached 79.2% with Claude Code + Opus 4.7 improving over ten continuous days of harness evolution.

We’re on Day 1 with Flash and started at 37% and are climbing.

Smaller model. New loop. Big bench.

---

Markdown version of https://cerebralvalley.ai/e/google-io-hackathon/hackathon/gallery/23. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt