One prompt, 3 hours, 22.8M tokens: local quantized model builds a GTA-style game
zmarcoz2 · reddit · 2026-10-01
Reddit user zmarcoz2 used a single prompt — "make a GTA-style game using three.js" — with a locally quantized qwen3.8-flash-next-iq3s model and a custom mini swe agent v2 harness. The run took 3h 18m and 22.8M tokens (22.5M input / 312K output) on an RTX 4080 Super 16GB, using the strata inference engine at 40 tok/s. The agent had powershell, file edit/read/search and image viewing tools, plus guards for tool failures and auto-compaction. Full logs are on a GitHub Gist.
More from coding & agent
- Dev says 70% of his workday runs on gpt-live inside Codex — tokenbender · 2026-10-01
- Asking Claude to simplify code spins up 8 parallel Opus agents and a million-token workflow — IgorBrigadir · 2026-10-01
- OpenBot launches open-source AI teammates for Mac running local models via Ollama — Robert-Prisacariu · 2026-10-01
- Developer admits he's addicted to coding with Claude Code on his phone — jdluk87 · 2026-10-01
- Runable launches autonomous cold outreach agent that finds leads and closes them 24/7 — SimplyAnnisa · 2026-10-01
- Git veteran warns: avoid SHA256 repos at Git 3 launch, they won't work for years — vmg · 2026-10-01