Read the memory label before buying hardware
Edge0’s published 35B result reports about 2.9 GiB of peak MLX allocation on a 24 GB M4 Pro Mac. That number describes one allocation measure inside a specific benchmark. It does not say the entire computer can run the model with three gigabytes of RAM. The operating system, file-backed weights, tokenization, caches and context can all need additional memory. Keep active allocation, total process memory and physical device RAM as separate measurements. We checked the project documentation and report; we have not independently reproduced the published benchmark.
Storage moves into the inference path
The engine keeps most model weights on storage and brings selected experts into active memory. The checkpoint therefore still has to fit on disk: the documented 35B checkpoint is roughly 23 GB, and the smaller 8B tier is about 4.2 GB. Storage capacity and read performance become practical constraints. A small memory headline alone does not establish the speed you will see on a different device or workload. Check the whole configuration before deciding that an existing laptop or phone is suitable.
A trained router predicts the next experts
The distinctive part of Edge0 is a trained prerouter. It predicts which experts will be needed next, allowing some reads to overlap the model’s computation. This is why the EDGE reel is a follow-up to a general storage-offload explainer. The model architecture, router and adapter files work together, so use the documented files for the chosen tier. The paper reports aggregate evaluation results; that is different from promising identical answers or byte-for-byte equivalence to a standard full-precision run.
ONE TACTIC A WEEK
Use the smaller model for your first run
On supported Apple Silicon hardware, the official Python route is a practical starting point. Read the current README, create an isolated Python environment and fetch the smaller model tier first. The companion setup guide provides the commands. Start with a short prompt and short output, then record load time, output speed, active allocation and process memory. Check that the generated answer is useful. Increase context gradually instead of treating one successful short response as evidence that every task will fit.
Mobile support has its own requirements
Android and iOS directories provide developer build routes. The Android documentation distinguishes the RAM requirements for 8B and 35B and asks for space for model sources and converted builds. The iOS workflow uses Xcode and model packs. These routes do not mean an app-store installation exists for every phone, or that the desktop benchmark applies unchanged to mobile hardware. Follow the platform-specific guide, check the device limits and record the same useful measurements you would collect on a Mac.
Keep cost, quality and compatibility separate
The engine is free Apache-2.0 code, but hardware, storage and model-specific terms still matter. The useful question is whether your own device produces acceptable answers at an acceptable speed and memory footprint. Retain the exact model tier and configuration with your results so a future change can be compared fairly. The EDGE guide links the source, technical report and both mobile workflows, and gives a short first-run checklist. It is a setup reference, not a claim that this benchmark has been reproduced on every supported platform.