A 35B MoE ran fully on a phone and shipped a Snake game in 10 minutes
dai_app · reddit · 2026-07-23
A developer says they ran a 35B MoE model entirely on a phone and used it to generate a Snake game in about 10 minutes. The app, Pocket Interpreter, lets the user describe an app, then writes and runs the code locally with no network access.
The post highlights that the generated Snake game included swipe controls, collision logic, score tracking, and speed ramping on the first try. It then argues that smaller models often fail on this kind of task, while bigger ones can get it right in one shot — which is exactly the challenge when trying to fit large models on mobile devices.
The broader technical point is the “BigMoeOnEdge” idea: a 120B MoE with only a few billion active parameters per token, where the always-needed parts stay resident and experts stream from flash, making a 12 GB phone a plausible host for a 60 GB model.
More from Infra
- Chinese open-source models may be reinforcing NVIDIA’s inference moat — zephyr_z9 · 2026-07-23
- AI portability is really about identity, traces, and rollback — not model IDs — Early-Matter-8123 · 2026-07-23
- OpenShip challenges AWS and Vercel with open-source deploys, MCP, and self-hosting — eyishazyer · 2026-07-23
- The Stack v3 expands to 114 TB with 224M repos and fixes a dedup bug — anton_lozhkov · 2026-07-23
- AI labs are now talking in gigawatts, and Europe is still planning in megawatts — AymericRoucher · 2026-07-23
- COLMAP merges DEGENSAC into main RANSAC, boosting plane-heavy scenes — ducha_aiki · 2026-07-23