MiniCPM5-2B Field Notes

Overview

Ahead of the MiniCPM5 release, OpenBMB reached out on X with an early run at the model and an open brief: trial it, show some use cases, have fun with it. The brief raised a more interesting question: how would people actually use a 2B model? Not in the cloud; on the phones and mobile devices people already carry.

So the field notes were built for the model's natural habitat: a catalogue of one hundred mobile UI proofs, each a single offline-first HTML file generated by the 2B model itself. The thinking behind it was simple design thinking: work backwards from where the model can actually run and who it serves. Each proof is a practical starting point, small enough to inspect and take further, for people who want to build their own custom apps on their own devices.

The guide is designed to be browsed cheaply: read the ideas first, spend bandwidth only on the build you choose. Every proof carries its own build record rather than an autoplaying demo, and the results were shared in the open on X.

The method

The point was never the artefacts themselves but the loop that produced them: measure how reliably a 2B model turns focused prompts into small, inspectable proofs through a repeatable generate-and-smoke-test cycle.

• Prompt: one capability, one mobile task
• Generate: one offline-first HTML artefact, 360px portrait
• Audit: load and exercise the observable UI
• Record: tokens, latency, attempts, outcomes

A passing record means the artefact survived the documented automated checks; it does not claim semantic correctness or production readiness. That honesty is the design.

Status

Live on GitHub Pages with all 100 build records.