Models · Chooser
Which model should I use?
Hebrish has several models, from tiny ones for wearables to a large one for desktop GPUs. Answer four questions and get one recommendation, with a runner-up.
From fastest to most accurate
Every model the chooser can pick, from the smallest and fastest on the left to the most accurate on the right. Your pick is green and the runner-up gold.
How it decides
- Live text means a streaming model: Shakshuka, or Sabich-asr4all-S on a wearable.
- Phones and computers without a GPU get Sabich, the model in the Hebrish apps. With speed first on a phone, the smaller Sabich-Parakeet-110M.
- An AMD or NVIDIA GPU with 6 GB of VRAM or more gets Sabich-Qwen3-ASR-1.7B (about 5 GB in bf16), unless speed comes first. With 4 GB or less, Sabich, or the Qwen model's 2 GB GGUF if accuracy comes first.
- Apple silicon gets Sabich's MLX build.
- Wearables and low-power boards get Sabich-Parakeet-110M, or Sabich-asr4all-S when speed comes first.
The rules and the positions above are in /data/variants.json. Every model is trained on one speaker's voice, so the gaps between them may be different for yours. Formats, versions and DOIs for each model: all models.