Spells 0.6.2

The first step towards Spells on Linux and macOS, and Show in folder finds every recording.

Added

• Diagnostics has a new System card. It names the operating system and lists anything Spells cannot do on it yet, which on Windows is nothing.

Changed

• Spells now reaches Windows through one small layer, the groundwork for versions for Linux and macOS. On Windows everything works as before.

Fixed

• Show in folder in History now selects the recording when a folder in its path has a space in its name. It opened the Documents folder instead.

Spells 0.6.1

Recordings you decide to include later reach your server too.

Fixed

• Switching on "Include the recordings" now also sends the recordings of dictations that were uploaded before without them. The server keeps one copy of each dictation and adds the recording to it.

Model files (models-v1)

The model files the Spells online installer downloads. You do not need to download anything from this page by hand: the installer picks the files your languages and your computer need and checks every one against the SHA-256 hash pinned in build/pins.json.

File	Size	What it does	Published by	Licence
Qwen3-ASR-0.6B-Q4_K_M.gguf	484 MB	Speech recognition for English and German	This project, see below	Apache-2.0
mmproj-Qwen3-ASR-0.6B-Q8_0.gguf	214 MB	Audio projector for Qwen3-ASR	ggml-org	Apache-2.0
ggml-large-v3-turbo-q8_0.bin	874 MB	Whisper large-v3 turbo, speech recognition for Albanian and every other language	ggerganov/whisper.cpp	MIT
ggml-silero-v5.1.2.bin	0.9 MB	Silero voice activity detection for Whisper	ggml-org	MIT
mtp-gemma-4-E2B-it-Q4_0.gguf	59 MB	Draft head that speeds up the Gemma 4 cleanup model	ggml-org	Apache-2.0

Qwen3-ASR-0.6B-Q4_K_M.gguf is this project's own quantisation of Qwen/Qwen3-ASR-0.6B: the bf16 weights converted to GGUF and quantised to Q4_K_M with llama.cpp b10997. Nothing else was changed. No one else publishes this file, so this release is its only source.

Two cleanup models are larger than the 2 GiB GitHub allows for one file, so the online installer downloads them from their publishers on Hugging Face, with the same hash check:

• gemma-4-E2B-it-Q4_0.gguf (2.8 GB, Apache-2.0) from ggml-org/gemma-4-E2B-it-GGUF
• Qwen3.5-4B-Q4_K_M.gguf (2.7 GB, Apache-2.0) from unsloth/Qwen3.5-4B-GGUF

The offline installer on the latest release carries every model it needs inside it, split into parts under 2 GiB.

The licence texts are attached as LICENSE-*.txt, and SHA256SUMS.txt lists the hash of every file here.