publicarray/fieldnote-models

Fieldnote models

Small language models 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 on the phone. A phone with a matching chip downloads a compiled bundle instead of compiling the model itself, which ran an iPhone 16 out of memory for the larger models.

Try Fieldnote: join the TestFlight beta (iPhone 15 Pro or later, iOS 27).

Bundles

Folder	Model	Chip	Phones	Licence
ios-h17g	MiniCPM5 1B	h17g	iPhone 16 family	Apache 2.0
ios-h17p	MiniCPM5 1B	h17p	iPhone 16 family	Apache 2.0
ios-h18p	MiniCPM5 1B	h18p	iPhone 17 Pro	Apache 2.0
minicpm5-2b/ios-h17p	MiniCPM5 2B	h17p	iPhone 16 family	Apache 2.0
minicpm5-2b/ios-h18p	MiniCPM5 2B	h18p	iPhone 17 Pro	Apache 2.0
qwen3.5-2b/ios-h17p	Qwen3.5 2B	h17p	iPhone 16 family	Apache 2.0
qwen3.5-2b/ios-h18p	Qwen3.5 2B	h18p	iPhone 17 Pro	Apache 2.0
lfm2.5-1.2b/ios-h17p	LFM2.5 1.2B Instruct	h17p	iPhone 16 family	LFM Open License 1.0
lfm2.5-1.2b/ios-h18p	LFM2.5 1.2B Instruct	h18p	iPhone 17 Pro	LFM Open License 1.0

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

Sources

Every bundle is a portable Core AI export compiled for one chip. No weights were changed.

Model	Original	Portable Core AI export, pinned revision
MiniCPM5 1B	openbmb/MiniCPM5-1B	mlboydaisuke/MiniCPM5-1B-CoreAI ios-static/ @ f38d143ec50cb2e54d8388479b919e3b9283943b
MiniCPM5 2B	openbmb/MiniCPM5-2B	mlboydaisuke/MiniCPM5-2B-CoreAI ios-static/ @ 39db5ff9480e5d80e423889e3bacb7f9e9b9e40f
Qwen3.5 2B	Qwen/Qwen3.5-2B	mlboydaisuke/qwen3.5-2B-CoreAI gpu-pipelined-b2/qwen3_5_2b_decode_int8hu_block32_sym @ 14e77014db5ae2f3de94825165f5ae58bededd7d
LFM2.5 1.2B	LiquidAI/LFM2.5-1.2B-Instruct	mlboydaisuke/LFM2.5-1.2B-CoreAI gpu-pipelined-b2/lfm2_5_1_2b_instruct_decode_int8hu_block32_sym @ 3e3241cbcf368a663768dcc4e5785c4f0a02bae4

Build

Compiled on GitHub Actions (Xcode 27, Metal Toolchain) by compile-models.yml:

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

The MiniCPM5 1B bundles leave the compute unit to Core AI. The Qwen3.5, LFM2.5 and MiniCPM5 2B h18p bundles are compiled for the Neural Engine. Some on-device specialisation still happens at first load, but not the full compile.

Loading

import CoreAI
import CoreAILanguageModels

let chip = AIModel.deviceArchitectureName           // e.g. "h17p"
let bundle = modelsDirectory.appending(path: "qwen3.5-2b/ios-\(chip)")
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.

Licences

This repo holds models under two licences. Each model's folder carries its own LICENSE:

• MiniCPM5 1B and 2B (OpenBMB) and Qwen3.5 2B (Qwen team, Alibaba Cloud): Apache License 2.0. The 1B's licence is the LICENSE at the repo root; the others are in minicpm5-2b/ and qwen3.5-2b/.
• LFM2.5 1.2B Instruct (Liquid AI): LFM Open License v1.0, in lfm2.5-1.2b/LICENSE. It allows commercial use only by organisations with under US$10 million in annual revenue; read it before using these bundles.

Core AI exports by mlboydaisuke; ahead-of-time compiled builds by publicarray. The files here are modified from the originals only by being compiled for a specific chip.