Local model finder.

We add the GGUF file’s weights, the model’s text attention cache and applicable recurrent state at your selected context, and a runtime allowance of 15% of the weights or 0.75 GiB, whichever is larger. Disk sizes use decimal GB. Memory estimates use GiB. We cap sliding-window caches where a separate window size is recorded. Otherwise the model uses a conservative full-length cache allowance.

We reserve at least 3 GiB or 20% of system memory for other apps, and at least 0.75 GiB or 10% of dedicated VRAM for display and drivers. Apple silicon shares one memory pool. Dedicated GPU and system memory are checked separately. On the DGX Spark, we instead apply a 120 GB decimal working-set ceiling (111.8 GiB) to its single 128 GB pool. Background apps and larger context windows reduce real headroom. See NVIDIA’s Spark runtime guide ↗.

Comfortable leaves at least 12% of the remaining budget free. Tight fit has less headroom. CPU only means the model fits system memory but exceeds available VRAM. CPU generation can be slower. Too large exceeds the estimated budget for a single pool. Mixed CPU/GPU offload may still work.

Auto selects 8-bit if it leaves at least 25% of the usable budget free. Otherwise it chooses the preferred available file in the best fit tier. Performance order uses published scores from one named test. Auto favors a general capability index when available for at least two matches, then the test with the most scored fits for the current task. Ties share the top score. The smaller working set appears first within a tie. These are estimates for one text session. Runtime, other apps, attention settings and prompt length affect actual usage. No laptop speed benchmarks are included.