Running a Plan-then-Execute Local Agent on 12GB VRAM: Big Model Plans, Small Model Acts
Loose_Doubt367 · reddit · 2026-10-31
Using an RX 6700 XT (12GB), R5 5600X, and llama.cpp with the Vulkan SDK, the author designs a local agent pipeline: a scheduled large model (Qwen3 27B, IQ3 quant) loads, plans tasks, saves its plan to disk, then unloads so a smaller model (Qwen3 35B A3B TieltCoder) executes the instructions for research, file management, or summarization. An optional loop lets the large model reload to verify results and issue new instructions on failure. The design philosophy: the small model should never re-verify—the large model front-loads all planning. The author is still in the planning phase and solicits model pairing suggestions.
More from coding & agent
- Codex 28-day challenge teases day-28 reset: OpenAI may be saving a new model for the finale — haider1 · 2026-10-31
- Rex Claude skill orchestrates multiple Claude Code sessions across Rex terminal panes — jasonkneen · 2026-10-31
- opencode Web UI bug: slash-command suggestions intermittently incomplete in long-running sessions — reisi007 · 2026-10-31
- ExpireMem Guardian gates AI code patches to catch removed APIs before they ship — Ok_Researcher391 · 2026-10-31
- Solo dev ships 3D tower defense game built with Claude and GPT images, now on Steam — MosskeepForest · 2026-10-31
- Curtis Yarvin proposes the "self-driving computer" where every app builds itself — ZeroStateReflex · 2026-10-31