CC-BY-NC-SA-4.0audioRuns locally3B
Hibiki-Zero: Specs, License & How to Run
Kyutai 3B streaming speech-to-speech (fr/es/pt/de→en). Weights CC-BY-NC-SA-4.0.
- License
- CC-BY-NC-SA-4.0
- Parameters
- 3B
- HF likes
- 59 (2026-10-27)
- First seen
- 2026-02-09
TL;DR. Hibiki-Zero is a audio model from Kyutai. Parameters: 3B.License: CC-BY-NC-SA-4.0. It is listed under DeepL, Otter.ai.
| Developer | Kyutai |
|---|---|
| Type | audio |
| Architecture | Decoder-only hierarchical Transformer (card: 3B backbone) that consumes and emits Mimi audio tokens plus text tokens at 12.5 Hz; audio bitrate 2.2 kbps. config.json model_type hibiki; dim 2048; 28 layers; 16 heads; Depth Transformer 6 layers / dim 1024. Same multistream family as Moshi / Hibiki. |
| Parameters | 3B |
| Modalities | audio, speech, text |
| Languages | fr, es, pt, de, en |
| License | CC-BY-NC-SA-4.0 — The lead card README states “License: CC BY-NC-SA 4.0” (opened 2026-10-27). Hugging Face cardData.license is empty and the card siblings list has no LICENSE blob. The inference repo kyutai-labs/hibiki-zero is MIT (GitHub license API + LICENSE file). Do not flatten weights and code into one SPDX. CC BY-NC-SA 4.0 is attribution + non-commercial + share-alike. Not commercial-free. Not legal advice. |
| Hardware | 8 GB (vendor README; 12 GB “safe”) — GitHub README: NVIDIA GPU; “8 GB VRAM should work, 12 GB is safe.” Card: batched inference 3× faster than real-time on a single H100 (vendor). HF usedStorage 6,648,922,558 bytes (~6.19 GiB). Peak VRAM is not independently measured. Python >=3.10,<3.14 (pyproject.toml 0.0.4). (not independently tested) |
| Hugging Face likes | 59 (as of 2026-10-27) |
| Model card | https://huggingface.co/kyutai/hibiki-zero-3b-pytorch-bf16 |
| First seen | 2026-02-09 |
Hibiki-Zero is Kyutai’s open simultaneous speech-to-speech checkpoint: a 3B hierarchical Transformer plus a hibiki-zero inference server. Official hosts we opened on 2026-10-27: kyutai/hibiki-zero-3b-pytorch-bf16 (the card the README and GitHub call “Hibiki-Zero”), github.com/kyutai-labs/hibiki-zero (278 stars; MIT), the blog Hibiki-Zero: Simultaneous Speech-to-Speech Translation Without Aligned Data (12 February 2026), and arXiv:2602.11072. This is Kyutai’s own lead card, not a community fine-tune.
This page hangs off DeepL alternatives (self-hosted translation search) and Otter.ai alternatives (live-speech search). Those mappings are editorial. DeepL is hosted text MT. Otter is a meeting-notes product. Hibiki-Zero is streaming speech-to-speech from French, Spanish, Portuguese, and German into English. We do not rank quality against DeepL, Otter, Interprefy, KUDO, or human interpreters.
Same-site speech peers: Hy-SimulMT (Tencent Hunyuan 30B-A3B simultaneous Chinese–English) — Hibiki-Zero vs Hy-SimulMT — and Granite Speech 5.0 (IBM file ASR/AST).
What the card says it does
Copied labels from the README we opened: simultaneous speech translation without word-level alignments; {French, Spanish, Portuguese, German}-to-English; real-time speech and timestamped text; voice transfer; temperature sampling (batch-friendly); trained on sequences up to 120 seconds; one speaker and one language per session, with a claimed zero-shot multi-speaker / mixed-language capability. Speech output is in-model (Mimi tokens) — not an optional external TTS service. Those sentences are upstream.
Hugging Face cardData.language is fr, es, pt, de, en. That matches the prose. The blog/paper’s Italian adaptation (<1000 h) is not a shipped pair on this card.
How it differs from Hy-SimulMT and Granite Speech 5.0
| Field | Hibiki-Zero | Hy-SimulMT | Granite Speech 5.0 |
|---|---|---|---|
| Job | Streaming S2ST + text | Streaming S2ST + text (SimulInfer) | File ASR / bidirectional AST |
| Pairs (prose) | fr/es/pt/de → en | zh ↔ en | Six-language ASR; AST to/from English |
| Size | 3B (card) | 30B-A3B | ~2.3B (5.0-2B) |
| Weights license | CC BY-NC-SA 4.0 (README) | Apache-2.0 | Apache-2.0 (card tag) |
| Voice | Built-in transfer (card) | Optional configured TTS | No speech-out on the 5.0-2B card |
This table is identification. We have no sourced bake-off. See Hibiki-Zero vs Hy-SimulMT and Hy-SimulMT vs Granite Speech 5.0.
What we verified (2026-10-27)
| Field | Value | Source |
|---|---|---|
| HF likes / downloads | 59 / 3,251 | HF API kyutai/hibiki-zero-3b-pytorch-bf16 |
| Card created / last modified | 2026-02-09T15:58:50Z / 2026-02-12T15:35:39Z | same |
| Gated | false | HF API |
| License (README body) | CC BY-NC-SA 4.0 | HF README |
| License (cardData) | empty | HF API |
| LICENSE blob on card | none | HF siblings |
| Pipeline tag | audio-to-audio | HF API / YAML |
| Config | model_type: hibiki; dim 2048; 28 layers |
HF config.json |
| usedStorage | 6,648,922,558 bytes (~6.19 GiB) | HF API |
| GitHub stars / forks | 278 / 25 | GitHub API |
| GitHub created / last push | 2026-02-09T15:29:05Z / 2026-09-09T14:36:54Z | same |
| GitHub license | MIT | GitHub API + LICENSE |
| Language | Python | GitHub API |
| Package version | 0.0.4 (>=3.10,<3.14) |
pyproject.toml |
| Blog date | 12 February 2026 | kyutai.org |
How to run (from official README, not yet tested)
The GitHub README’s one-command path:
uvx -p 3.13 hibiki-zero serveThen open the URL it prints. --gradio-tunnel is documented for a public tunnel. Without uv: pip install hibiki-zero then hibiki-zero serve. File / batch path: hibiki-zero generate --file /path/to/audio.wav. NVIDIA GPU required. forkalt has not run those commands.
License
Weights: the card README we opened says CC BY-NC-SA 4.0. Hugging Face cardData.license was empty. There is no LICENSE file among the six card siblings. Read CC BY-NC-SA 4.0.
Code: kyutai-labs/hibiki-zero/LICENSE is MIT (“Copyright (c) 2026 kyutai”). GitHub’s license API agrees.
Do not treat MIT inference code as a commercial grant on the 3B weights. Not legal advice.
What we did not verify
Vendor BLEU / ASR-BLEU / speaker-similarity / average-lag rows on the Kyutai blog stay self-reported. Subjective audio-quality tables stay vendor. The H100 “3× faster than real-time” sentence is upstream. Peak VRAM, live lag on our hardware, and Italian as a shipped pair are not filled in with guesses.
Related: Hy-SimulMT — Hibiki-Zero vs Hy-SimulMT, Granite Speech 5.0 — Hy-SimulMT vs Granite Speech 5.0, Nemotron 3.5 Transcribe (file ASR, not streaming S2ST) — Nemotron 3.5 Transcribe vs Granite Speech 5.0, LibreTranslate (text MT API), whisper.cpp, DeepL alternatives, Otter.ai alternatives.
Benchmarks
Quoted from the cited source. forkalt has not independently reproduced these numbers.
| Name | Score | Source |
|---|---|---|
| Fr→En BLEU (vendor blog) | 30.6 (self-reported) | https://kyutai.org/blog/2026-02-12-hibiki-zero/ |
| Fr→En avg. lag seconds (vendor blog) | 6.1 (self-reported) | https://kyutai.org/blog/2026-02-12-hibiki-zero/ |
| Es→En BLEU (vendor blog) | 32.3 (self-reported) | https://kyutai.org/blog/2026-02-12-hibiki-zero/ |
Deploy options
- local: Hugging Face weights (3B BF16)
- local: hibiki-zero serve (GitHub)
- hosted: Official samples space
Comparisons
FAQ
Is Hibiki-Zero an official Kyutai model?
Yes. The ungated lead card is huggingface.co/kyutai/hibiki-zero-3b-pytorch-bf16 (created 2026-02-09T15:58:50Z). Inference code is github.com/kyutai-labs/hibiki-zero (created 2026-02-09T15:29:05Z). The 12 February 2026 Kyutai blog names the same ids. It is not a third-party fine-tune.
Are the weights open for commercial use?
The card README says CC BY-NC-SA 4.0. That is a non-commercial, share-alike grant. Hugging Face cardData.license was empty in the 2026-10-27 API check and the card has no LICENSE file. The GitHub inference repo is MIT. Do not treat the MIT code license as a weight grant. Not legal advice.
Does it replace DeepL, Otter, or Hy-SimulMT?
No. DeepL is hosted text MT. Otter is a meeting-notes product. Hibiki-Zero is a local streaming speech-to-speech checkpoint (French, Spanish, Portuguese, German → English) with built-in voice transfer. Hy-SimulMT is Tencent Hunyuan’s Apache-2.0 30B-A3B simultaneous Chinese–English stack. Different pairs, size, and license. We have no sourced bake-off.
Which languages are actually shipped?
Card and blog: French, Spanish, Portuguese, and German into English. One speaker and one language per session. The blog/paper mention an Italian adaptation with less than 1000 hours as a research result — Italian is not in Hugging Face cardData.language and is not listed as a shipped pair here.
Sources & verification
- https://huggingface.co/kyutai/hibiki-zero-3b-pytorch-bf16
- https://github.com/kyutai-labs/hibiki-zero
- https://github.com/kyutai-labs/hibiki-zero/blob/main/LICENSE
- https://kyutai.org/blog/2026-02-12-hibiki-zero/
- https://arxiv.org/abs/2602.11072
- https://creativecommons.org/licenses/by-nc-sa/4.0/
License notes are not legal advice. Read the original LICENSE. Trademark & disclaimer.