The Race — Building Democracy in the Age of AI. People race toward a shared finish line.
The Race
← Trends

Supporting case

MiniCPM's deployment package makes a larger model a local mobile component

What happened

OpenBMB released MiniCPM-V 4.7 around October 12–13, with available Apache-2.0 weights and a mobile deployment repository ↗. The 35B-A3B version contains roughly 35 billion parameters while activating about three billion per generated token. Its image/video-capable model uses “experts,” subsets of model weights selected for different computations; the runtime streams them from device storage instead of keeping all of them in working memory.

The package includes local iOS, Android and HarmonyOS routes and source code. Its recommended 35B mobile configuration requires about 17 GB of free storage and at least 12 GB device RAM. The publisher's headline 13-plus tokens/second and roughly 3 GB working-memory claim is not a 3 GB total download or a third-party-verified result. Storage speed, quantization, device setup and application integration remain dependencies.

An available local runtime can widen hardware alternatives. It does not establish that an entire organization’s AI workflows are independent.

October 28, 2026 — document-processing release

OpenBMB announced MiniCPM-OCR on October 28. Its released parser turns scanned documents into structured text, including Markdown or JSON. A separate model, MiniCPM-OCR-Jev, compares an image region with a proposed transcription and returns a quality score. That score can help send questionable passages to a human reviewer; it does not guarantee that the text is correct. Code and models are offered under Apache 2.0. Primary repository ↗; parser weights ↗; checking model ↗.

The project documents local deployment routes, including common local-model runtimes. Its fuller document-processing workflow also needs layout analysis and software integration. Published performance examples use specified hardware; they do not establish the same speed on a typical office computer. Accuracy and throughput claims remain developer evidence, rather than an independent operational evaluation.

For organisations handling sensitive documents, a usable local parser can reduce reliance on an external document API. The quality-checking component offers a way to target human attention, but is not an independent accountability institution.

Dates and implementation stage

Earlier mobile deployment released October12–13; document-processing release October28, 2026.

Supporting sources

Related trends