Where the Bits Go: Sensitivity-Allocated Mixed Precision for a 4B Reasoning Model on a 12 GB Consumer GPU

On consumer GPUs, reasoning language models are constrained chiefly by memory capacity and bandwidth rather than by compute. For Qwen3.5-4B, a hybrid of Gated DeltaNet and full-attention layers, we measure how sensitive each class of tensors is to quantization and spend a fixed size budget where the measured damage per byte is largest, choosing only among standard GGUF formats, so that the result needs no training and runs unmodified in Ollama. Within the range of weight sizes we tested (about 1.4–2.0 GiB), fewer bytes placed by this allocation beat more bytes placed at random: at each of three budgets the allocation diverges less from BF16 than all three random placements of the same bytes, and the allocation at 1.80 GiB diverges less than every random placement at 2.04 GiB. Against a uniform assignment of types the evidence is weaker: at the smallest budget the two tie, the allocation using 0.04 GiB less, and at the middle budget the allocation is about 1 point of divergence lower by interpolation. The resulting model, magi-1.0, served with an 8-bit KV cache, reduces resident GPU memory 3.4-fold relative to BF16 (3.0-fold with the default 16-bit cache), retains 96% of BF16's index on a 12-suite benchmark while its error rate rises from 16.0% to 19.4%, decodes 2.7 times as fast and achieves 3.0 times the capability per gibibyte. Relative to a published 2-bit file of similar size, its advantage comes almost entirely from fewer degenerate reasoning loops. A standard 4-bit file remains more accurate than this configuration by 3.5 index points but requires 1.13 GiB more memory, so magi-1.0 is preferable only when memory is the binding constraint.