KaniTTS to Gepard: which model to use

There are 28 nineninesix repositories on Hugging Face. Most people arrive at one of them from a search result and have no idea whether they have landed on the current model, an old one, a backbone, or a checkpoint from the middle of a training run.

This is the map. It covers what each generation actually changed, what the naming convention means, and the licence difference that matters more than any of the benchmark numbers.

What are the three generations?#

KaniTTS, KaniTTS2 and Gepard. Each generation replaced the backbone, and the last one also replaced the licence.

KaniTTS (370M / 400M). A two-stage pipeline: a language model predicts audio tokens, an FSQ neural codec turns them into a waveform. The backbone is Liquid AI's LFM2. This generation went widest on languages, with per-language fine-tunes for English, German, Arabic, Chinese, Spanish, Korean and Kyrgyz, plus MLX ports so it runs on Apple Silicon.

KaniTTS2 (400M). Same two-stage shape, with frame-level position encodings and tuning aimed squarely at realtime conversation. It is the most liked model we have published — kani-tts-2-en has 184 likes against Gepard's 131 — which is a reasonable proxy for how well it landed with people running models locally.

Gepard 1.0 (555.7M). A different design goal. Where Kani was a good model that could be served, Gepard is a model shaped by the constraint that it must run on stock vLLM without patching anyone's compute kernels. The backbone is Qwen3.5: 14 blocks, hidden dimension 1024, 8 attention heads, about 500M parameters, plus an audio interface and a voice-cloning compressor to reach 555.7M. Audio is NVIDIA NeMo NanoCodec, FSQ, 22.05 kHz, 21.5 frames per second, 1.89 kbps.

Which model do people actually download?#

Downloads from the Hugging Face API, read on 25 August 2026:

The second row is worth explaining, because 2,197 people downloading a backbone is not 2,197 people using a TTS model. qwen3_5-full-attn-only-14 is the full-attention Qwen3.5 variant Gepard is built on. It is published because reproducing our training requires it, not because it does anything on its own.

What do the suffixes mean?#

The naming is systematic once you know the rule, and confusing until you do.

So kani-tts-450m-0.2-pt is a pretrained base at the 450M size, version 0.2, and kani-tts-450m-0.2-ft is what came out of fine-tuning it. If you are looking for something to run rather than something to train from, you want a -ft or a language-suffixed model, not a -pt.

Which licence covers which weights?#

This is the part that costs people time if they get it wrong, so it goes in its own section rather than a footnote.

Gepard's audio codec carries its own terms as well: NanoCodec is under the NVIDIA Open Model License Agreement, separate from the Apache 2.0 on our weights.

If your legal review turns on this, the longer piece on open weights against closed vendors works through what Apache 2.0 does and does not grant.

How does Gepard score, including where it loses?#

Measured on the public Seed-TTS-eval set: 1,088 paired prompts, identical UUIDs and texts across every model.

Read the row honestly. Gepard leads on every perceived-quality axis: highest naturalness on NISQA-MOS, cleanest on noise, colouration and discontinuity. It is sixth of seven on word error rate and last on speaker similarity by a wide margin, 0.585 against 0.867 for the leader.

That is the trade the streaming-first design bought. If your product needs a cloned voice to be mistakable for the original, SIM 0.585 is a real problem and you should look elsewhere. If it needs a clean, natural voice that starts speaking in 50 milliseconds, the same row reads as a good deal.

Where does Diamond fit?#

It does not, quite. diamond-1.0 is in the same organisation and is not a TTS model at all: it restores degraded speech to 44.1 kHz using an autoregressive RQ-Transformer over Descript Audio Codec tokens. 166.6M parameters, trained from scratch in 63 GPU-hours, Apache 2.0.

Mentioned here only because it shows up in searches next to the speech models and gets mistaken for one.

So which one do I use?#

For anything shipping to users today the answer is Gepard 1.0, and the rest of this list exists because a handful of real cases point elsewhere. All 28 repositories are on the nineninesix Hugging Face organisation.

• Voice agent, production, English or Spanish or Portuguese or Dutch. Gepard 1.0, either self-hosted or through the hosted API. The API is $5 per million characters.
• Running locally on a Mac. KaniTTS MLX ports. Gepard's realtime figures assume a CUDA GPU.
• A language Gepard does not cover. A KaniTTS language fine-tune, or fine-tune Gepard yourself.
• Training your own from our base. qwen3_5-full-attn-only-14 plus the gepard-train pipeline.
• Cleaning up bad recordings. Diamond, and it is unrelated to everything above.