Mistral Medium 3.5 Crawls on Strix Halo

This is a useful reality check: a 128B dense open-weight model can be impressive and still feel unusable once you push it onto consumer-class hardware with a huge context load.

• –The bottleneck is not just the model size, but the combination of 128B dense weights, long context, and local inference overhead
• –Strix Halo’s unified memory makes this possible, but not fast enough for interactive codebase work
• –The numbers suggest Mistral Medium 3.5 is better suited to batch jobs, offline analysis, and queued agent runs than live back-and-forth prompting
• –For developers, the practical takeaway is that “self-hostable” and “pleasant to use” are still very different bars
• –The post also reinforces why quantization and server flags matter less than raw architecture when the model is this large