Muesli models are on Hugging Face
Download Muesli dictation cleanup models in MLX, GGUF, and LiteRT formats, plus CoreML speech recognition ports for Apple Silicon.
We are publishing the models we fine-tune and port for Muesli on Muesli HQ on Hugging Face. Download the weights, read the prompts, and use them in your own local speech tools under each model’s license.
Our work covers dictation cleanup and speech recognition on Apple Silicon. Publishing it lets other developers inspect our choices, reproduce the documented tests, and build on the same models.
Dictation cleanup models
The cleanup models take a transcript and are trained to remove fillers, apply spoken corrections, format dictated lists, and fix clear recognition mistakes while preserving the speaker’s wording. They support English and optional app context.
- Qwen3.5 0.8B: MLX 8-bit for Apple Silicon, with a GGUF version for llama.cpp.
- Qwen3.5 2B: MLX 8-bit and GGUF builds for the larger Qwen cleanup model.
- LFM2.5 1.2B: MLX 4-bit, MLX 8-bit, and GGUF options.
- Gemma 4 E2B: MLX 4-bit for Apple Silicon and a LiteRT-LM build.
The cards include training details, exact prompt formats, and evaluation results. Use those prompts when testing. The reported cleanup benchmarks use authored examples; performance on your own dictations can differ.
Speech recognition ports for Apple Silicon
- Indic-Transcribe Flex CoreML supports mixed-script output, including Hindi in Devanagari alongside English in Latin script.
- Indic-Transcribe Core CoreML produces native-script output for Indic speech.
The Indic ports have been checked against upstream transcripts in a small set of smoke tests. Their tested configuration uses CPU and GPU on macOS 15 or later. Broader language coverage and long recordings need further evaluation.
Build with us
These releases build on Qwen, Liquid AI, Google, Bodhan AI / AI4Bharat, and NVIDIA. Each card identifies its upstream model and license. Check those terms before reuse, especially the custom licenses attached to LFM and Indic models.
Start with the Muesli model collection, or explore the Muesli source code. For the app, download Muesli. Our Bodhan CoreML and MLX guide goes deeper into the Apple ports.