Local AI Coding on RTX 3060 12GB: Qwen2.5-Coder 14B Works but Tool Use Breaks
GoalHefty3604 · reddit · 2026-11-19
A user set up a local AI coding server with Ollama on a headless Linux box (RTX 3060 12GB, 32GB DDR4) and tested qwen2.5-coder:14b with Qwen Code from another PC.
- Context capped at 32K with constant compacting; faking a 128K window in Qwen Code stopped the compacting but broke tool use — the model asked for copy-paste or emitted raw JSON instead of calling tools
- With the Continue extension in VS Code, the 1.5b model handled autocomplete but felt dumb; 14b chat worked but had the same tool-use issues
- Author's verdict: low quality and quirky, asking whether this hardware can do better
More from coding & agent
- Snowflake Labs open-sources Cortex AI 101 notebook calling OpenAI models via SQL — adnan_hashmi · 2026-11-19
- Agents spent years hillclimbmaxxing while the human at the keyboard became an afterthought — tokenbender · 2026-11-19
- Google's GUARD paper: frozen base model, self-editing agent harness gains +8.89 points — omarsar0 · 2026-11-19
- Microsoft open-sources openclaw-dev: one-command hosted OpenClaw agent on Azure with sandboxed tool execution — adnan_hashmi · 2026-11-19
- 6-minute Hubble documentary made with Claude and Jimeng CLI workflows — xiaohu · 2026-11-19
- Coding agents gone wild: test-cheating, scope drift, and rogue DB migrations — Common_Dream9420 · 2026-11-19