Same Qwen3-Coder 30B: instant success via LM Studio, 57-minute fix loop via Ollama
Proof_Nothing_7711 · reddit · 2026-10-11
An Italian developer tested Qwen3-Coder-30B-A3B (UD-Q8KXL) with Cline on a Ryzen 9 + W7800 48GB eGPU, using the same prompt to build a local photo-upscaling webapp. Via Cline + LM Studio (Vulkan or ROCm), the agent shipped a working 56-71KB app in 250-310 seconds on the first try. Via Cline + Ollama, the agent spiraled into 120-290MB React/nodemodules builds with broken UI, failing to fix itself even after 57 minutes. Takeaway: the inference stack dramatically changes agent behavior; the model itself was consistently good at JS audits and bug fixes.
More from coding & agent
- Meta & CMU's IdeaScientist uses RL agents for cross-domain research ideation, lifting novelty from 36.3% to 67.0% — ZeYanjie · 2026-10-11
- Agents on a trading MCP server backtest 6x more than humans but almost never deploy — QuanTradin · 2026-10-11
- A notes MCP server that lets Claude Code, Codex and Gemini share one notebook — lovegrover · 2026-10-11
- Google's TabFM: a zero-shot foundation model that predicts table data in one forward pass — Prompt Engineering · 2026-10-11
- O'Reilly 'Agent Memory' book enters early release with first two chapters live — danielrock · 2026-10-11
- saccade: A Local-First Python Library for Video Transcripts, Frames, and Semantic Search — Dapper_Ad599 · 2026-10-11