Summary
Fahd Mirza tests Nanbeige 4.2, the second release from Nan Beach and a 3-billion-parameter model built on a “looped transformer” architecture that challenges conventional scaling assumptions. Where standard transformers stack separate layers — each adding parameters with every depth increase — Nanbeige 4.2 runs a single shared-weight layer through multiple iterative passes, refining its understanding with each loop without adding parameters. The result, according to vendor benchmarks, is a 3B model that outperforms 9B and 12B models from competing families including Qwen 3 3.6B and Gemma 4 12B across most agent and reasoning benchmarks, with the exception of a narrow gap on SideCode where Gemma 4 12B holds a slight edge.
The video walks through the complete local installation on Ubuntu with an Nvidia RTX 3060 (48 GB VRAM), including a critical caveat: Nanbeige 4.2 requires a custom-patched vLLM fork that is not in upstream vLLM, meaning the build must compile entirely from source — a process Mirza clocks at roughly four hours. Once served, the model consumes approximately 44 GB of VRAM. The functional test assigns it a FastAPI report generator with a Redis pub/sub bug (publisher and subscriber using mismatched channel names, causing jobs to queue indefinitely) via a Hermes agent with no hints. The model eventually finds and fixes the bug, but runs for over 43 minutes and surfaces numerous additional issues beyond the single planted defect — a behavior Mirza attributes to the unusually long chain-of-thought the looped architecture generates even at 4B total parameters.
📺 Source: Fahd Mirza · Published July 23, 2026
🏷️ Format: Hands On Build
