Model card and licence

b6ea3b5 verified 29 days ago

2.46 kB

license: apache-2.0
base_model: openbmb/MiniCPM5-1B
pipeline_tag: text-generation
library_name: coreai
tags:
  - core-ai
  - ios
  - on-device
  - minicpm5

Fieldnote models

MiniCPM5 1B compiled ahead of time for Apple's Core AI, one ready-to-load bundle per iPhone chip family. Used by Fieldnote, an on-device meeting recorder, to write meeting notes without compiling the model on the phone (the on-phone compile of a larger model ran an iPhone 16 out of memory).

Bundles

Folder	Core AI architecture	Phones
`ios-h17g`	h17g	iPhone 16 family
`ios-h17p`	h17p	iPhone 16 family
`ios-h18p`	h18p	iPhone 17 Pro

Each folder is a complete bundle: metadata.json (pointing at the compiled model), tokenizer/, and minicpm5_1b_minicpm5_pal8_g32_static.<arch>.aimodelc. A phone needs only the folder whose architecture matches AIModel.deviceArchitectureName. Other devices should use the portable model below and let Core AI compile it.

Source and build

Model: openbmb/MiniCPM5-1B, 8-bit palettized, 4,096-token context.
Portable Core AI export: mlboydaisuke/MiniCPM5-1B-CoreAI ios-static/ at revision f38d143ec50cb2e54d8388479b919e3b9283943b.
Compiled on GitHub Actions (Xcode 27, Metal Toolchain) by compile-models.yml:

xcrun coreai-build compile minicpm5_1b_minicpm5_pal8_g32_static.aimodel \
  --platform iOS --min-deployment-version 27.0 --architecture <arch> --output <dir>

Compute units were left to Core AI's default choice. Some on-device specialization still happens at first load, but not the full compile.

Loading

import CoreAI
import CoreAILanguageModels

let arch = AIModel.deviceArchitectureName           // e.g. "h17g"
let bundle = modelsDirectory.appending(path: "ios-\(arch)")
let model = try await CoreAILanguageModel(resourcesAt: bundle)

Fieldnote downloads only the matching folder, from a pinned revision of this repo, and checks every file against its SHA-256 before use.

Licence

MiniCPM5 is released by OpenBMB under the Apache License 2.0, and so are these compiled bundles (see LICENSE). No weights were changed; the files here are the portable export compiled for specific chips.