Mistral Medium 3.5 benchmarked on 3x3090

This is less a launch story than a deployment reality check: Mistral says Medium 3.5 can self-host on as few as four GPUs, and this post shows what that looks like in the wild.

• –Q3 quantization is doing the heavy lifting here; without it, a 128B model would be much harder to fit into 72GB VRAM.
• –The different prompt types matter: code, SVG, and HTML stress the model and runtime differently, so the screenshots are more useful than a single synthetic tokens/sec number.
• –For local AI builders, this is the core tradeoff: you get an open-weight frontier-ish model, but you still pay in GPU count, power, and tuning effort.
• –If the quality holds up at this speed, Medium 3.5 becomes interesting for self-hosted coding agents, not just chat demos.