Ferrous-7B: a mixture-of-experts model that runs on a laptop

Only 1.8B parameters active per token. The quantized build fits in 6GB of VRAM and matches last year's 13B dense models on code tasks.

Summarized by the Neuricon wire. Every claim links back to its primary source where available.