Hibiki-Zero vs Hy-SimulMT

Official Hibiki-Zero Temporal Transformer and Depth Transformer architecture figure from the Kyutai paper — Kyutai Hibiki-Zero paper, Temporal Transformer / Depth Transformer figure. Source · License: Kyutai paper figure (arXiv:2602.11072).

Official Hy-SimulMT RealSI GEMBA versus StreamLAAL figure from the Hugging Face model card — Tencent Hunyuan Hy-SimulMT-30B-A3B card (RealSI GEMBA / StreamLAAL figure). Source · License: Apache-2.0.

Verdict. Pick Hibiki-Zero if you want Kyutai’s ungated 3B streaming speech-to-speech stack for French, Spanish, Portuguese, or German into English (built-in voice transfer; weights CC-BY-NC-SA-4.0). Pick Hy-SimulMT if you want Tencent Hunyuan’s Apache-2.0 30B-A3B simultaneous Chinese–English stack (SimulInfer; optional TTS). Same job class, different pairs and licenses. Vendor tables stay self-reported. We have no sourced bake-off.

Compare Hibiki-Zero and Hy-SimulMT.

Dimension	Hibiki-Zero	Hy-SimulMT	Source
License (weights)	CC BY-NC-SA 4.0 (README; cardData.license empty; no LICENSE blob)	Apache-2.0 (card tag + LICENSE-Hy-SimulMT.txt)	Hugging Face / GitHub (2026-10-27)
License (inference code)	MIT (kyutai-labs/hibiki-zero)	Apache-2.0 (Tencent-Hunyuan/Hy-SimulMT LICENSE file)	GitHub LICENSE files
Gating	No	No	Hugging Face API
Developer	Kyutai	Tencent Hunyuan	Official cards
Parameters	3B (card); usedStorage ~6.19 GiB	30B-A3B; HF safetensors 30,394,794,496	Official cards / HF API
Job	Streaming S2ST + timestamped text (fr/es/pt/de → en)	Streaming S2ST (zh ↔ en prose) via SimulInfer	Official cards
Voice / speech-out	In-model Mimi tokens + voice transfer	Speech-out and cloning need a configured TTS service	Official cards
Serving named	hibiki-zero serve / generate (uvx or pip)	SimulInfer + vLLM >=0.31	Official READMEs
Hugging Face likes	59 (2026-10-27)	18 (2026-10-27)	Hugging Face API
Quality vs DeepL / Otter	Not independently tested	Not independently tested	n/a

When to choose which

Live French / Spanish / Portuguese / German → English: Hibiki-Zero — Card prose is those four pairs into English. Not a quality rank.
Live Chinese–English simultaneous interpretation: Hy-SimulMT — Card prose is bidirectional zh–en. Hibiki-Zero does not list Chinese.
Apache-2.0 / commercial-friendly weights: Hy-SimulMT — Hibiki-Zero weights are CC-BY-NC-SA-4.0. Not legal advice.
Compact ~3B local S2ST: Hibiki-Zero — Hy-SimulMT is 30B-A3B (~56.7 GiB BF16). Size is not a quality claim.

Read the entity pages: Hibiki-Zero and Hy-SimulMT. Both hang off DeepL alternatives and Otter.ai alternatives. File ASR/AST on this site is Granite Speech 5.0 — Hy-SimulMT vs Granite Speech 5.0.

Why this comparison exists

Searchers who want open simultaneous speech translation now have Kyutai’s 3B Hibiki-Zero next to Hunyuan’s 30B-A3B Hy-SimulMT. Same job class (streaming S2ST). Different language pairs, licenses, and serve paths. This is identification, not a bake-off.

Sourced snapshot (2026-10-27)

Field	Hibiki-Zero 3B	Hy-SimulMT-30B-A3B
Host	huggingface.co/kyutai/hibiki-zero-3b-pytorch-bf16 (not gated)	huggingface.co/tencent/Hy-SimulMT-30B-A3B (not gated)
HF likes	59	18
Weights license	CC BY-NC-SA 4.0 (README)	Apache-2.0 (LICENSE file + card tag)
Size	3B card; ~6.19 GiB usedStorage	~30.39B safetensors; 30B-A3B name
Pairs (prose)	fr/es/pt/de → en	zh ↔ en
Serve path	`hibiki-zero serve`	SimulInfer + vLLM >=0.31
Quality vs DeepL / Otter	Not independently tested	Not independently tested

License

Hibiki-Zero’s weights README says CC BY-NC-SA 4.0; cardData.license was empty and the card had no LICENSE blob. The code repo is MIT. Hy-SimulMT’s GitHub file says “Hy-SimulMT is licensed under Apache-2.0.” Not legal advice.

How to choose

Need live French / Spanish / Portuguese / German → English and accept a non-commercial weight grant: Hibiki-Zero.
Need live Chinese–English simultaneous speech-to-speech with Apache-2.0 weights: Hy-SimulMT.
Need file ASR/AST or speaker tags: Granite Speech 5.0.
Need text MT (not speech): LibreTranslate on DeepL alternatives.
Need local Whisper STT: whisper.cpp.

FAQ

Which is better at translation quality?

We have no sourced bake-off. Hibiki-Zero’s BLEU / lag rows and Hy-SimulMT’s FLEURS / RealSI / GEMBA rows stay vendor. Do not merge those tables.

Can I use either as an Otter or DeepL clone?

No. Neither page is a meeting-notes product or a hosted text MT API. Both stream speech translation. See Otter.ai alternatives and DeepL alternatives.

Why this pairing instead of Granite Speech 5.0?

Both are simultaneous speech-to-speech pages on this site. Granite Speech 5.0 is file ASR/AST — already compared with Hy-SimulMT. The clearer split here is language pair + license + size.

Are the licenses the same?

No. Hibiki-Zero weights are CC BY-NC-SA 4.0 on the card README; the inference repo is MIT. Hy-SimulMT is Apache-2.0 on the card and LICENSE-Hy-SimulMT.txt. Not legal advice.

Sources & verification

Last verified 2026-10-27 · Last updated 2026-10-27· Edited by forkalt editors · Editorial policy

License notes are not legal advice. Read the original LICENSE. Trademark & disclaimer.