Acknowledgements

Patum is built on open models and code. Thank you to their authors. The full license texts are in the app, under Help › Acknowledgements.

Models

Qwen3 4B Instruct 2507

Summaries, briefings and Live context answers on this Mac. Apache-2.0. Copyright Alibaba Cloud (Qwen team). 4-bit MLX conversion by the MLX Community (mlx-community/Qwen3-4B-Instruct-2507-4bit).

Qwen3.5 9B

Summaries and briefings on Macs with more memory. Apache-2.0. Copyright Alibaba Cloud (Qwen team). 4-bit MLX conversion by the MLX Community (mlx-community/Qwen3.5-9B-MLX-4bit).

Qwen3 Embedding 0.6B

Live context search. Apache-2.0. Copyright Alibaba Cloud (Qwen team). 4-bit MLX conversion by the MLX Community (mlx-community/Qwen3-Embedding-0.6B-4bit-DWQ).

BGE Reranker v2 M3

Live context search ranking. Apache-2.0. By the Beijing Academy of Artificial Intelligence (BAAI). Converted to a smaller MLX format by the Patum authors.

Whisper large-v3 turbo

Transcription (Balanced). MIT. Whisper by OpenAI. Core ML conversion by Argmax (argmaxinc/whisperkit-coreml, MIT). Tokenizer files from openai/whisper-large-v3 on Hugging Face (Apache-2.0).

Parakeet TDT 0.6B v3

Transcription (Faster). CC-BY-4.0. By NVIDIA. Modified: converted to Core ML by Fluid Inference (FluidInference/parakeet-tdt-0.6b-v3-coreml).

Silero VAD

Finding speech in recordings. MIT. By the Silero Team. Core ML conversion by Fluid Inference (FluidInference/silero-vad-coreml, MIT).

NVIDIA Nemotron 3 Diarization

Speaker labels: who spoke when. OpenMDW-1.1. Nemotron 3 Diarization was released by NVIDIA under the OpenMDW License Agreement, version 1.1. Core ML conversion, unchanged here, by Fluid Inference (https://huggingface.co/FluidInference/nemotron-3-diarization-coreml).

WeSpeaker ResNet34-LM

Speaker labels: linking voices across a meeting. CC-BY-4.0. WeSpeaker (https://github.com/wenet-e2e/wespeaker), trained on VoxCeleb, as packaged by pyannote. Modified: converted to Core ML by the Patum authors.

pyannote community-1 PLDA

Speaker labels: linking voices across a meeting. CC-BY-4.0. PLDA of pyannote's speaker-diarization-community-1; the parameters originate from BUT Speech@FIT (VBx) and are distributed by Fluid Inference (FluidInference/speaker-diarization-coreml). Modified: transforms folded into one file by the Patum authors.

NVIDIA TitaNet-L

Speaker labels: telling voices apart and recognizing known people. CC-BY-4.0. NVIDIA NeMo's titanet_large speaker verification model, from the ONNX export distributed by k2-fsa/sherpa-onnx. Modified: rebuilt in PyTorch and converted to Core ML by the Patum authors.

Ported code

DiariZen

The VBx clustering setup and PLDA transforms after Nemotron 3 (ported to Swift). MIT.

VBx

Speaker clustering (ported to Swift). Apache-2.0. By L. Burget and P. Pálka, BUT Speech@FIT. Rewritten in Swift with changes (double precision, the GMM path only, PLDA decomposition at export time).

pyannote.audio

Diarization reconstruction and embedding masks (ported to Swift). MIT.

ebur128

EBU R128 loudness measurement in the audio tools and tests (ported to Swift; itself a port of libebur128). MIT.

BIP-39 English word list

The team root's 24 words and the six words read when joining a team. MIT.

silero-rs

Voice activity state machine (ported to Swift). MIT.

Swift packages

EventSource

Streamed responses (through swift-huggingface). MIT.

FluidAudio

Core ML speech models: Nemotron 3, Parakeet, Silero VAD. Apache-2.0.

GRDB.swift

The meeting database (SQLite). MIT.

MLX Swift

Local models on the GPU. MIT.

MLX Swift LM

Language models, embeddings and the reranker on MLX. MIT.

Sparkle

Updates outside the App Store (the edition sold from the site). MIT.

SQLCipher

Encrypts local databases. BSD-3-Clause.

Swift Collections

Data structures. Apache-2.0.

Swift Crypto

Hashing (through swift-transformers). Apache-2.0.

Swift Hugging Face

Model files and tokenizers. Apache-2.0.

Swift Jinja

Chat templates. Apache-2.0.

Swift Numerics

Complex numbers for MLX. Apache-2.0.

Swift Transformers

Tokenizers. Apache-2.0.

WhisperKit

Whisper speech recognition on Core ML. MIT.

yyjson

JSON parsing (through swift-transformers). MIT.