Model licenses

Last updated: 5 September 2026

Weeve runs its AI features on your own device. To do that, it downloads machine-learning models to that device: the Weeve macOS app (the “App”) keeps them on your Mac, and the Weeve for Chrome browser extension (the “Extension”) keeps them in your browser profile. Those models are made by third parties and licensed by them, not by Weeve.

This page lists every model Weeve downloads, who provides it, and the terms it comes under. For how Weeve handles your data, see our Privacy Policy.

Weeve does not bundle model weights in the App or in the Extension, and no model weights are included in the download from our website or from the Chrome Web Store. The models below are fetched when you first use the feature that needs them, and they stay on your device. The App downloads them from Hugging Face. The Extension downloads them from Weeve’s own content delivery network at models.getweeve.io.

The App and the Extension do not use the same models. Each list below applies only to the product it names.

Models the macOS app downloads

Summaries: Gemma 4

Provided by Google DeepMind. Weeve offers Gemma 4 in several sizes so it can match how much memory your Mac has; every size is covered by the same licence. The builds Weeve downloads were converted and quantised for Apple silicon by the mlx-community project, from Google’s released weights.

Licensed under the Apache License, Version 2.0. Google publishes the model license at ai.google.dev/gemma/docs/gemma_4_license.

Transcription: Parakeet TDT 0.6b v3

Provided by NVIDIA. Weeve downloads a version converted to run on Apple silicon by FluidInference (FluidInference/parakeet-tdt-0.6b-v3-coreml). The conversion changes the format the model runs in; it does not retrain the model.

Licensed under Creative Commons Attribution 4.0 International. Attribution required by that licence: “Parakeet TDT 0.6B v3” by NVIDIA Corporation, huggingface.co/nvidia/parakeet-tdt-0.6b-v3, licensed under CC BY 4.0. Changes were made: converted to Core ML by FluidInference.

Speaker labels: pyannote community-1

Provided by the pyannote authors. Weeve downloads a version converted to run on Apple silicon by FluidInference (FluidInference/speaker-diarization-coreml). As above, the conversion changes the format, not the model itself.

Licensed under Creative Commons Attribution 4.0 International.

This model works out who spoke when. It builds on pyannote segmentation, WeSpeaker speaker embeddings, and VBx clustering. The FluidAudio SDK that runs it is a separate piece of software, licensed under the Apache License, Version 2.0.

Models the Chrome extension downloads

The Extension runs in a browser, so it uses ONNX and MLC builds of these models rather than the Apple silicon builds the App uses. All of them are served from models.getweeve.io. The same list, with the full licence texts, ships inside the Extension and is opened from Settings, under About, with the Licenses row.

Transcription: Parakeet TDT 0.6B v3 (ONNX)

Provided by NVIDIA. The Extension uses an ONNX export of that model published by ysdede (ysdede/parakeet-tdt-0.6b-v3-onnx). Weeve serves that export in two precisions: a 16-bit encoder for browsers with WebGPU and an 8-bit encoder for the WebAssembly fallback, with an 8-bit decoder in both cases. Those are format and precision changes only; the model was not retrained.

Licensed under Creative Commons Attribution 4.0 International. Attribution required by that licence: “Parakeet TDT 0.6B v3” by NVIDIA Corporation, huggingface.co/nvidia/parakeet-tdt-0.6b-v3, licensed under CC BY 4.0. Changes were made: exported to ONNX by ysdede (huggingface.co/ysdede/parakeet-tdt-0.6b-v3-onnx), and quantised to fp16 and int8 as described above. The full legal code is at creativecommons.org/licenses/by/4.0/legalcode.

Speaker segmentation: pyannote segmentation 3.0

Provided by the pyannote authors (pyannote.audio, CNRS), huggingface.co/pyannote/segmentation-3.0. The Extension uses the ONNX conversion distributed by the sherpa-onnx project. This model works out when each speaker is talking.

Licensed under the MIT License.

Speaker embedding: 3D-Speaker CAM++

Provided by the 3D-Speaker project (Alibaba Group), github.com/modelscope/3D-Speaker. The Extension uses the campplus-sv-en-voxceleb-16k model in the ONNX conversion distributed by the sherpa-onnx project. This model tells voices apart so segments can be grouped per speaker.

Licensed under the Apache License, Version 2.0.

To do’s: Weeve Nano (Qwen3-0.6B)

Weeve Nano is our name for the small language model that drafts to do’s from a transcript. The weights are Qwen3-0.6B, provided by Alibaba Cloud (huggingface.co/Qwen/Qwen3-0.6B). The Extension uses the MLC q4f16_1 conversion (mlc-ai/Qwen3-0.6B-q4f16_1-MLC), which changes the format and precision only. It needs WebGPU, and it is downloaded on request rather than with the other models.

Licensed under the Apache License, Version 2.0. The MLC runtime and the compiled model library that runs it are separate software, also under the Apache License, Version 2.0.

Previously distributed models

The App has changed which summary models it offers over time. If you installed an earlier version, you may still have one of these on your Mac, and Weeve will move you to a current model automatically:

• Earlier Gemma 4 builds, under the Apache License, Version 2.0, as above.

• Qwen 3.5 2B (mlx-community/Qwen3.5-2B-4bit), provided by Alibaba Cloud and licensed under the Apache License, Version 2.0.

Removing models

In the App, remove any downloaded model at any time from Settings > Models, or remove all of them from Settings > Data & Storage. In the Extension, remove them from Settings, under Models. Removing the Extension from Chrome deletes its copies as well.

Changes to this page

We update this page when the models Weeve uses change. The “Last updated” date above reflects the most recent change.

Questions: support@getweeve.io