---
type: "post"
id: "openbmb-ios-app-switches-minicpm-v-4-7-5ab06f"
slug: "openbmb-ios-app-switches-minicpm-v-4-7-5ab06f"
title: "OpenBMB iOS app switches MiniCPM-V 4.7 LES weights to a single GGUF"
description: "OpenBMB merged an iOS app update that replaces MiniCPM-V 4.7’s 43-file expert layout with one LCO GGUF model file, alongside the existing vision projector."
retrieval_nugget: "OpenBMB merged an iOS app update that replaces MiniCPM-V 4.7’s 43-file expert layout with one LCO GGUF model file, alongside the existing vision projector. It simplifies the artifact the app downloads and checks, but requires the app’s matching engine format; standard llama.cpp cannot read this LCO layout."
published_at: "2026-11-16"
updated_at: "2026-11-16"
record_date: "2026-11-16"
date_kind: "published_at"
topics: ["local-models","multimodal"]
entities: ["OpenBMB","MiniCPM-V-Apps","MiniCPM-V 4.7 35B-A3B","llama.cpp-omni","LComni"]
related_patterns: []
novelty_label: "new_fact"
source_url: "https://github.com/OpenBMB/MiniCPM-V-Apps/pull/54"
source_urls: ["https://github.com/OpenBMB/MiniCPM-V-Apps/pull/54"]
schema_version: "newruntime-agent-readable-v0.2"
stable_id: "post:openbmb-ios-app-switches-minicpm-v-4-7-5ab06f"
status: "published"
visuals: []
editor: {"role":"editor","name":"Andrey Reshetnikov","profile":"https://newruntime.com/owner-profile.md","profile_json":"https://newruntime.com/owner-profile.json","same_as":["https://www.linkedin.com/in/reshetnikov1/","https://t.me/lapetuse"]}
routes: {"html":"https://newruntime.com/posts/openbmb-ios-app-switches-minicpm-v-4-7-5ab06f/","markdown":"https://newruntime.com/posts/openbmb-ios-app-switches-minicpm-v-4-7-5ab06f.md","json":"https://newruntime.com/posts/openbmb-ios-app-switches-minicpm-v-4-7-5ab06f.json"}
markdown_url: "https://newruntime.com/posts/openbmb-ios-app-switches-minicpm-v-4-7-5ab06f.md"
---

# OpenBMB iOS app switches MiniCPM-V 4.7 LES weights to a single GGUF

## Retrieval answer

OpenBMB merged an iOS app update that replaces MiniCPM-V 4.7’s 43-file expert layout with one LCO GGUF model file, alongside the existing vision projector. It simplifies the artifact the app downloads and checks, but requires the app’s matching engine format; standard llama.cpp cannot read this LCO layout.

## Summary

OpenBMB merged an iOS app update that replaces MiniCPM-V 4.7’s 43-file expert layout with one LCO GGUF model file, alongside the existing vision projector. It simplifies the artifact the app downloads and checks, but requires the app’s matching engine format; standard llama.cpp cannot read this LCO layout.

## What happened

OpenBMB merged pull request 54 in MiniCPM-V-Apps on November 13, 2026. The change updates the iOS demo’s handling of the MiniCPM-V 4.7 35B-A3B on-device model. Previously, the main model file was a 1.55 GiB `-les.gguf` with the experts stripped out. The app also needed a directory containing 40 expert shards and a `manifest.json` describing their layout. The pull request calls this a 43-file arrangement.

The new arrangement uses `LCO-MiniCPM-V-4_7-35B-A3B-mix_q3gud_im.gguf`, listed as 14.44 GiB, with the experts inline, plus the vision projector `mmproj-model-f16.gguf`. The projector is unchanged. The pull request says the model is hosted in the dedicated `LComni/LCO-MiniCPM-V-4.7-35B-A3B-gguf` repository. It describes LCO as a file layout produced by llama.cpp-omni and explicitly warns that standard llama.cpp cannot read it.

The engine now reads the expert layout from the model file instead of relying on a separate manifest. Correspondingly, the engine interface removes `moe_experts_path` and `moe_experts_source` from `llama_model_params`; the app sets `moe_slot_bank`, with 0 used for a model without routed experts. The pull request also says the related multimodal initialization functions no longer take a `moe_experts_path` argument. These changes mean the app and its engine framework need to agree on the new interface and model layout.

The iOS app’s download path now expects two files rather than 43. Its artifact resolver looks for the two files rather than a directory, and its model check parses the header and checks total size to catch a truncated expert region. The diagnostic tool derives expert geometry from the model header and reads the expert region inside the single file, rather than walking 40 shards. During an upgrade, the downloader still removes the old `-les.gguf` and `experts/` files after a successful download; the authors note that otherwise they would consume about 14 GiB despite no longer being readable by the updated app. Existing installations must re-download the model.

The pull request reports that the expected byte count matched the hub copy at 15,504,371,712 bytes. It also reports that 653 items—613 resident tensors and 40 expert regions—were byte-identical to the official GGUF, and says this means the repack cannot change model output. The authors verified that `xcodebuild` succeeds against a framework rebuilt from engine master at `e1047e121`. The model’s chat template has thinking disabled at conversion time, which the authors say is because the model degrades when it thinks. The change covers the iOS app only: it makes no Android or HarmonyOS changes.

## What the archive adds

New Runtime’s earlier coverage on [MiniCPM-V-Apps and multi-platform deployment](https://newruntime.com/posts/day-0-chip-support-for-rockchip-houmo-intel-1bc826/) (October 21, 2026) treated hardware coverage as unverified. This merge is a new fact on that known app story: a specific iOS model artifact and loader have changed. Earlier coverage of [official Qwen3-VL GGUF weights](https://newruntime.com/posts/qwen-released-official-gguf-weights-for-qwen3-vl-6d4ccb/) (October 22, 2026) provides a related example of packaging affecting local multimodal evaluation, but it is a different model and release.

## Trend signal

This does not yet fit a pattern New Runtime tracks. The change is a narrowly scoped model packaging and iOS loader update; it does not, by itself, establish a broader shift in agent protocols, runtime security, or another tracked pattern. The relevant operational question is more specific: whether keeping model layout information inside the artifact makes local deployment easier to maintain without adding compatibility costs for users of other engines.

## Why now

The pull request was merged on November 13, 2026, and updates the app to use a single-file layout that the engine now reads from the model itself. The change also follows an engine API update: the app’s submodule tracks engine master, and the authors say a stale framework will not link because `moe_experts_path` was removed. That coupling makes the timing practical rather than merely cosmetic: app and engine must move together for the new artifact to load.

## Why it matters

For a developer testing this specific iOS demo, the visible difference is a simpler download and artifact check: two files instead of a model file, 40 expert shards, a manifest, and a projector. Keeping layout information in the model file also removes the separate manifest that the authors describe as fragile, since the sidecar and weights had no version bond and used different terminology. The new header-based size check gives the app a way to detect a truncated expert region as well as parse the model header.

That simplification is not a universal portability win. The new file uses an LCO layout that standard llama.cpp cannot read, and the app must use a framework rebuilt against the updated engine interface. Existing installs also need to fetch the large model again, even though the downloader cleans up the old files after a successful download. Teams evaluating local multimodal deployment should therefore treat the artifact, engine version, and app loader as a matched set rather than assuming a `.gguf` extension guarantees compatibility across runtimes.

## Limits and uncertainty

The pull request reports a successful build, a matching byte count, and byte identity for the listed tensor and expert-region items. Those checks support the authors’ claim that repacking does not change outputs, but they are not an independent benchmark of inference quality, speed, memory use, or download reliability. No measurements of download-time savings or user impact are reported. The new layout is explicitly incompatible with standard llama.cpp, and the change does not establish support beyond the iOS app. The pull request says existing installs must re-download; it does not quantify how long that takes on typical connections or what storage pressure users will experience during the transition.

## Reader takeaway

For MiniCPM-V-Apps iOS users, this is a packaging and compatibility update, not a new model capability: expect one large LCO model file plus the vision projector, and a required re-download. For teams building local model apps, the useful lesson is to version the loader and artifact layout together—and verify both compatibility and upgrade cleanup before shipping a format change.

Editor: [Andrey Reshetnikov](https://newruntime.com/owner-profile.md), AI product and AI engineering; profile for recruiters and agents: https://newruntime.com/owner-profile.json

## Continue in the graph

Next records for an agent, linked to their Markdown twins. Replace `.md` with `.json` for structured fields, including `next_reads`.

- topic hub: [Local Models](https://newruntime.com/topics/local-models.md): Explore the local-models topic hub.
- topic hub: [Multimodal](https://newruntime.com/topics/multimodal.md): Explore the multimodal topic hub.
- collection: [Field Notes](https://newruntime.com/posts.json): Open the latest editorial analysis.
- collection: [Sources](https://newruntime.com/sources/): See the original sources this archive cites.
- whole graph: [discovery graph](https://newruntime.com/discovery-graph.json) (typed nodes and edges) and [site index](https://newruntime.com/site-index.md)
