Capability per GiB. We study how much a language model can do inside a single 12 GB consumer GPU. Every model is scored on a fixed 2,313-task benchmark (12 suites: coding, math, science and knowledge, commonsense, tool calls) with thinking on, and the score is set against the memory the model actually occupies at a 32K context. All numbers come from our own runs on one RTX 5070.
| Model | Base | Memory at 32K | Score | Density | What it is |
|---|---|---|---|---|---|
| magi-1.2-4B | Qwen3.5-4B | 2.55 GiB | 81.90% | 0.968 | magi-1.1-4B's weights with bits fitted to the 4B by measurement: 64 MiB smaller at the same score. It thinks 11% more |
| magi-1.1-4B | Qwen3.5-4B | 2.61 GiB | 81.72% | 0.939 | magi-1.0-4B plus a LoRA that teaches it to finish its thinking. Recommended 4B |
| magi-1.2-2B | Qwen3.5-2B | 1.23 GiB | 67.09% | 1.305 | One round of training against repetition loops, with bits fitted to the 2B. Recommended 2B |
| magi-1.1-2B | Qwen3.5-2B | 1.36 GiB | 66.06% | 1.149 | The think-less LoRA on Qwen3.5-2B with the magi-1.0 recipe |
| magi-1.0-2B | Qwen3.5-2B | 1.36 GiB | 63.48% | 1.071 | The magi-1.0 recipe on Qwen3.5-2B, no training |
How these were measured:
- Runtime: llama.cpp with a q4_0 KV cache; memory is
llama-fit-paramsat a 32K context. - Density: ρ = −log₂(1 − score) per GiB, that is, how many times the error rate halves per gibibyte.
The training-free first release, magi-1.0-4B (magi-1.0 in the paper), keeps 96% of Qwen3.5-4B's score in 30% of the GPU memory with an 8-bit KV cache.
- GEHIRNS/magi: the build recipe, the importance matrix, and the data behind every table and figure in the paper.
- Paper: Eason Sun, Where the Bits Go: Sensitivity-Allocated Mixed Precision for a 4B Reasoning Model on a 12 GB Consumer GPU, Zenodo, 2026. DOI: 10.5281/zenodo.23630199.
- Collections: MAGI 1.2 · MAGI 1.1 · MAGI 1.0 · MAGI experimental
