Blogger runs blind test of five self-hosted AI agents on local hardware

A user ran a blind test of five self-hosted agents (OpenCode, goose, Atomic Agent, Hermes, OpenClaw) on an M5 Max MacBook Pro, all using the same quantized Qwen 3.8 27B model in llama.cpp across five everyday tasks. All passed contact merging and log error counting; four passed bug fixing and receipt-to-spreadsheet, where two installed a missing Python library unprompted and one failed after asking for permission. Total times ranged from about 10 to 24 minutes.

• KoboldCpp ships built-in agent harness2026-09-26
• Community discusses Qwen4 expectations and model size needs2026-10-26
• Qwen4-27B may feature 4-way parallel residual streams2026-10-26
• Qwen3.8 Flash Next hits 262 tok/s on 16GB VRAM2026-10-25
• Nearly half of new Korean AI models build on Chinese foundations2026-10-25