Models · NVIDIA · Parakeet TDT-CTC 110M
Sabich-Parakeet-110M small devices
All of Sabich's training on NVIDIA's smallest Parakeet, about a fifth of the size. For low-end phones, watches and wearables.
At a glance
| Base model | nvidia/parakeet-tdt_ctc-110m (NVIDIA) |
| Current version | Sabich-Parakeet-110M v1.0, released 2026-10-21 |
| Parameters | 114M (base model card) |
| Evaluation | Not benchmarked yet Model card |
| Download | every format and quant (Hugging Face) |
| Collection | Hebrish/sabich on Hugging Face |
| Licence | CC-BY-4.0 · use governed by the terms of use |
| Used by | Your own tools (sherpa-onnx, transcribe.cpp); not in the Hebrish apps yet |
| DOIs | none yet |
Formats and quants
| Repository | Format and files | Version | For | DOI |
|---|---|---|---|---|
Hebrish/Sabich-Parakeet-110M main | .nemo (NeMo) .nemo 459 MB, training config · access request · files | v1.0 | NVIDIA NeMo | none yet |
Hebrish/Sabich-Parakeet-110M-sherpa-onnx-int8 | sherpa-onnx int8 encoder, decoder, joiner, about 137 MB · access request · files | v1.0 | Android, Linux, Windows, small boards | none yet |
Hebrish/Sabich-Parakeet-110M-GGUF | GGUF Q8_0 135 MB, Q4_K_M 90 MB, F16 229 MB · access request · files | v1.0 | transcribe.cpp (Offline Voice Input, Handy) | none yet |
Hugging Face issues a DOI per repository, so each format can be cited on its own. Repositories, sizes and DOIs checked on the Hub on 2026-10-21.
Versions
| Version | Released | Run | Where | Notes |
|---|---|---|---|---|
| Sabich-Parakeet-110M v1.0 current | 2026-10-21 | r8-ctc110m-all | Sabich-Parakeet-110M | Sabich's data (blend all, revision e918e381): 592.4 min of the speaker's takes (382.9 min code-switching) plus 66.0 min LibriSpeech replay; 20 epochs on one L40S |
About Sabich-Parakeet-110M
Sabich-Parakeet-110M v1.0 (2026-10-21) is Sabich, the everything model, trained on nvidia/parakeet-tdt_ctc-110m: a FastConformer encoder with a TDT decoder and an extra CTC head, 114M parameters. Same data as the other Sabich variants: code-switching in the canonical romanization, tech vocabulary, brand names, acronyms, numbers and currency in digits, and everyday email and message dictation. Output is cased and punctuated.
Not benchmarked yet. It was published straight after training. Its validation WER at the chosen checkpoint was 9.43% (raw, on the speaker's validation takes, used only to pick the checkpoint): not a test score, and not comparable across base models. Expect it to be less accurate than the 0.6B Sabich; it is the choice when size and speed matter more.
Run r8-ctc110m-all: dataset revision e918e381, 592.4 min of the speaker's takes (382.9 min code-switching) plus 66.0 min of LibriSpeech-PC replay, 5,634 clips, 20 epochs on one L40S (40 minutes); best checkpoint epoch 17.
Run it
- sherpa-onnx int8 (about 137 MB): the same kind of build the Hebrish apps run, for Android, Linux, Windows and small boards.
- GGUF for transcribe.cpp, which Offline Voice Input on Android and Handy on the desktop use: take
Sabich-Parakeet-110M-Q8_0.gguf(135 MB). - NeMo, from the .nemo file, like any Parakeet model.
- It is not in the Hebrish apps yet.
Cite
Sabich-Parakeet-110M has no DOI yet; cite the repository, Hebrish/Sabich-Parakeet-110M.
@misc{rosehill_sabich_110m_2026,
author = {Daniel Rosehill},
title = {Sabich-Parakeet-110M},
year = {2026},
publisher = {Hugging Face},
url = {https://huggingface.co/Hebrish/Sabich-Parakeet-110M}
}Request access
Request access on Hugging Face first. The Sabich-Parakeet-110M repositories are gated: open the repository, sign in, and fill in the short request form there. Requests are reviewed by hand. For non-commercial use, the attestation form covers what is asked.
No answer after a few days? Request access here instead, or write to access-requests@hebrish.app. Commercial use, including telephony, needs a commercial licence.
Which model should I use? · All Sabich variants · All Hebrish models