Qwen3.5 hits limits on local rigs

The core lesson is that “local AI” is mostly a hardware-and-format problem before it is a model-selection problem. Qwen3.5 is strong, but 9B-class models can still feel punishing on CPU-only boxes, and Jetson-class devices need very careful model sizing, quantization, and software compatibility.

• –For CPU-only inference, smaller quantized coder models will usually beat a larger “best quality” model that loads slowly or fails outright.
• –`failed to read magic` usually points to a bad download, the wrong file format, split-file confusion, or an older/incompatible `llama.cpp` build, not just a random runtime crash.
• –Jetson Nano 4GB is extremely tight for modern 9B models; even if a model technically loads, practical throughput and memory pressure can make it unusable.
• –A Tesla P40 would help on the DX380 if the chassis, power, cooling, and PCIe constraints can be solved, but it will not fix format or loader issues.
• –The practical path is to standardize on a current `llama.cpp` build, use a verified GGUF quantization from a trusted source, and benchmark smaller coder models before chasing a larger one.