Edge0 open-sources 145B-parameter and 120B on-device models with inference framework and app, claiming laptop operation at 29GB peak memory

Edge0 has published open weights for two models, Edge0-145B and Edge0-120B, along with an inference framework, deployment tools and a desktop app. The 145B model is a sparse mixture-of-experts design with about 10B parameters active per step. Edge0 says it runs on a compatible laptop at 29GB peak memory, with 1,100+ tokens/s prefill, 80+ tokens/s decode and about 120W power draw. It reaches this through shared-matrix experts, three-tier loading across GPU, RAM and SSD, and speculative decoding. All figures come from Edge0's own testing, and the extraction found no independent verification. It also does not say which laptop counts as "compatible".

If the numbers hold on ordinary hardware, developers and small firms could run a large coding and agent model locally. That would remove per-token fees, keep data on the machine and work offline, which matters most for privacy-sensitive enterprises. The open weights and framework mean outsiders can test the claims directly, which is a point in its favour. Until they do, the real impact is uncertain: a laptop that fits 29GB peak memory is a high-spec machine, not a typical one, and decode speed under long contexts or sustained load is unreported.

"World's first open-source on-device 145B model" is a packaging claim that depends on how "on-device" is defined, since a sparse model with 10B active parameters is not comparable to a dense 145B one. "Rivalling frontier cloud models" rests on self-reported benchmarks and should not be taken at face value. Missing context includes the exact hardware, quantisation level, context length, and how output quality changes once experts are loaded from SSD.

1. 01Independent reproductions of the 29GB peak memory and 80+ tokens/s decode figures on named consumer laptops, including long-context and sustained-load runs, within the next few weeks.
2. 02Third-party benchmark results (not Edge0's own) on coding and agent tasks against the models Edge0 compares itself with, such as DeepSeek-V4-Flash-0731 and Qwen3.8-Flash-Next.
3. 03Adoption signals: whether the framework and shared-matrix expert approach are picked up by other runtimes and projects, and whether GitHub activity, issues and the licence terms show real use rather than a one-off launch.