Can Edge0-145B run locally? It comes down to whether the memory math adds up: the official figure of 29GB peak memory refers to the GPU side, not to a whole machine needing only 29GB of memory. This 145B-parameter MoE model uses about 10B parameters per inference step, and the official claim is that it can run on a compatible laptop — but before running it, add up the graphics card, system memory, and storage figures, and check the official installation guide to confirm your system and graphics card are supported.
The 29GB Figure Is Really Three Separate Bills
| Where | Official figure | What it holds |
|---|---|---|
| GPU side | 29GB peak | The most VRAM-hungry part of inference |
| System memory | About 40GB | The expert library, loaded into the GPU on demand |
| SSD | About 48GB | Vocabulary data, read on demand |
The official blog calls this approach a "four-layer design": expert matrices are shared across expert groups, and the CPU and GPU compute together. If you buy a machine based on 29GB alone, the math will be wrong.
Check Your Machine in Three Steps
- Check the graphics card and operating system first: the official blog uses a compatible laptop as its example, but the supported range is defined by the official installation guide — match your system and GPU model against it before going further.
- Then add up memory and storage: on top of the 29GB GPU peak, you also need about 40GB of system memory and about 48GB of SSD space. All three are required.
- Finally, confirm where to get it: the model weights are on Hugging Face, and the inference framework and Edge0 App are on GitHub. Read the licenses before downloading — the inference framework is Apache-2.0, the model weights use the Edge0 Community Model License 1.0, and you must read both licenses carefully before any commercial use.
peak memory, decode speed of 80+ tokens/s, a 128K–256K context window, and the average benchmark score of 72.8 across six tests are all Edge0's own internal test figures, with no independent third-party reproduction yet — don't treat them as results you will get out of the box. The official blog also says local inference has no cloud token or API fees and data stays on your machine, but you still pay for the hardware and electricity yourself. Typical power draw is stated at about 120W, so a laptop's cooling and power supply count too — running a large model is not just about the memory number.
