LISL: a lightweight agentic loop claiming 1.2 tasks per token on a local Gemma 12B
MammothTreat6751 · reddit · 2026-11-12
The author shares LISL, a lightweight implementation-style agentic loop, claiming "1.2 tasks per token." The accompanying screenshot shows LM Studio running Gemma 4 12B QAT locally: a local REST API server with tool calling, a system prompt casting the model as a "GBA game architect" designing turn-based tactical games, logs showing 2.2 tok/s inference and 20k+ token task runs.
It's essentially a demo of running a minimal agent loop on consumer local models, though the post itself gives almost no implementation detail.
More from coding & agent
- Leak: xAI Working on a Selection Box for Grok Bot to Comment on Files — nima_owji · 2026-11-12
- Coding Agent Tooling Layers: One Agent vs Many, Personal vs Org-Wide — jiayuan_jy · 2026-11-12
- MobAI: Rotate Driver/Navigator Roles in AI Group Work to Boost Team Autonomy — JoeJustice · 2026-11-12
- Running a personal AI agent for months: architecture, failures, and open questions — Powermadlux · 2026-11-12
- Fable 5.2 and Opus 5.5 rebuild a low-poly cat in Blender with exact triangle budgets — FinanceYF5 · 2026-11-12
- freellmapi-setup-cli: Interactive CLI validates LLM keys and builds budget-balanced model chains — biryaniwithachaar · 2026-11-12