The wise cheese
Local ML voice chat using high-end models, aiming for a self contained, user-friendly out-of-the-box experience as much as possible.
This is a work in progress with frequent updates; TestFlight builds are available here for macOS, iOS and visionOS.
| Selection | Full | Mini |
|---|---|---|
 |
 |
 |
A light helper app which can run on an iPhone or device with not enough processing power, which automatically detects and connects to Emeltal on the network and provides the same voice interface. Testflight link for this app is here
Emeltal offers a curated list of proven open-source high-performance models, aiming to provide the best model for each category/size combination. This list often changes as new models become available, or others are superceeded by much better performing ones. Most models (with the exception of certain extremely large variants, which are capped at 16384 tokens) run at their maximum context size.
- [Qwen 2.5 72b] (https://huggingface.co/bartowski/Qwen2.5-72B-Instruct-GGUF)
- [Qwen 2.5 32b] (https://huggingface.co/bartowski/Qwen2.5-32B-Instruct-GGUF)
- [Qwen 2.5 14b] (https://huggingface.co/Qwen/Qwen2.5-14B-Instruct-GGUF)
- [Qwen 2.5 7b] (https://huggingface.co/bartowski/Qwen2.5-7B-Instruct-GGUF)
- Dolphin 2.9.2 on Qwen 2.5
- Dolphin 2.7 on Mixtral
- Dolphin 2.9.3 on Mistral Nemo
- Dolphin 2.8.1 on TinyLlama
- Emeltal heavily relies on the llama.cpp for LLM processing, and whisper.cpp for voice recognition.
- Text rendering uses Ink to convert between Markdown and HTML.
- Uses my PopTimer for debouncing things.
Released under the terms of the MIT license, see the LICENSE file for license rights and limitations (MIT).
All model data which is downloaded locally by the app comes from HuggingFace, and use of the models and data is subject to the respective license of each specific model.
Copyright (c) 2023-2024 Paul Tsochantaris