---
title: "Pick mobile models by RAM class: notes from a Shenzhen bench"
url: https://memory.wiki/hB8Nii1t
updated: 2026-11-14T21:58:48.421Z
source: "memory.wiki"
---
# Pick mobile models by RAM class: notes from a Shenzhen bench

## Notes on "Pick mobile models by RAM class"

Leave headroom for the OS.  Local AI is still a hands-on field. Many of the useful answers come from people who share their exact configuration, what failed and how they fixed it, rather than from polished announcements.

## What makes a useful post

- Full specs: CPU, GPU, RAM, storage, operating system, driver and runtime versions.
- The model file and quantisation, with where it came from.
- What you tried, in order, including the failures.
- Measured observations with the method described, never numbers copied from elsewhere.
- Noise, heat and power observations if the box lives in a shared room.

## Offline and mobile use

Running small models on phones, tablets or mini PCs is practical for note-taking, translation drafts and offline FAQs. Pick model size by available RAM and leave room for the operating system. Expect heat during long sessions; short prompts and pauses help. For kiosks and offline demos, keep content signed and updates controlled.

## Helping newcomers

A new user benefits from a single path: one runtime, one small model, one evaluation set and a checklist. Avoid sending beginners through five tools at once. Encourage them to write down what they did; their notes often become the next person's tutorial.

## Keeping a failure wiki

Organise entries by symptom, configuration, fix, owner and date. Titles should match what people search for, such as the exact error message. Review old entries when runtimes change, and mark the ones that no longer apply.

## Ways to contribute

- Reproduce someone else's setup and report whether it worked for you.
- Turn a long troubleshooting thread into a short summary with the final fix.
- Translate a useful guide and credit the original author.
- Add measurement methods to posts that only show results.
- Flag outdated tutorials when a runtime or model changes.

Small contributions add up. A community that documents its failures honestly saves everyone time and keeps beginners from repeating the same mistakes.

More on this topic: [DIY lab](https://microaione.com/diy/). See also [local AI tutorials](https://microaione.com/tutorials/).

## Bottom line

Local AI rewards patience and record keeping more than raw spending. Decide what the system must do, measure it on your own hardware and documents, write down every change with a date and an owner, and review the setup after each significant upgrade. Small, documented steps are easier to debug, easier to hand over and easier to defend when someone asks why a particular choice was made. When in doubt, test on a smaller scale first and expand only when the evidence supports it.

## Quick reminders

- Log the model name and version with every request.
- Pin driver and runtime versions together.
- Review the setup again after every major model upgrade.
- Leave memory headroom for the OS and the embedding model.

*Published by [MicroAI One (本地AI 1号网)](https://microaione.com/), a Shenzhen team working on local LLM hardware, software and deployment.*

---

## Summary
Successful local AI implementation relies on choosing model sizes based on available RAM while leaving headroom for the operating system. Documenting specific configurations, failures, and testing methods creates a reliable knowledge base that simplifies debugging and future upgrades.

## Themes
- Local AI hardware optimization
- Documenting troubleshooting processes
- Mobile model deployment
- Community knowledge sharing

## Key takeaways
- Always leave RAM headroom for the operating system when selecting a model size.
- Useful troubleshooting posts must include full hardware specs, driver versions, and specific model quantization details.
- New users should follow a single path consisting of one runtime, one model, and one evaluation set to avoid confusion.
- Failure wikis should be organized by symptom, configuration, and fix to facilitate easier searching.
- Local AI setups should be reviewed and re-evaluated after every major model or runtime upgrade.

## Insights
- Successful local AI deployment relies more on meticulous record keeping and iterative testing than on high end hardware investment.
- Failure documentation should be structured by symptoms and specific configurations rather than just successful outcomes to prevent repetitive errors.
- Mobile AI performance is heavily constrained by OS memory overhead, requiring users to prioritize headroom over model size.

## Open questions / gaps
- What specific metrics or tools are recommended for measuring heat and power consumption in mobile devices?

