AnalysisAI ModelsSeptember 20, 2026

Qwen 3.8 27B runs 21-day local agent loop on one RTX 3090

A Reddit user ran a local agent loop for ~21 days on a single RTX 3090 with Qwen 3.8 27B, tasking it to build a CUDA inference engine for its own GPU architecture. It produced working kernels and benchmarks but no win over llama.cpp; compaction consumed ~83 hours across ~12 human messages.

1 source

More stories today

Open the live feed