KittenTTS 2
A 1.7B ternary text-to-speech model for CPUs, with 115 voices, voice cloning and emotion tags. Stellon Labs Community License.
In short. KittenTTS 2 is a text-to-speech model small enough to run on an ordinary computer's processor. It has 115 built-in voices, can copy a voice from a short recording, and lets you add emotions and sounds such as laughs with simple tags. It is free for research and personal use, and free for businesses until they pass US$1 million in revenue or funding.
Summary
KittenML released KittenTTS 2 around October 2026. We give only the month because its Hugging Face repository was created on 30 September, while the public GitHub activity we could see for it runs from 21 to 27 October. It is a 1.7-billion-parameter speech language model that reads text and writes S3 codec tokens, which a vocoder turns into 24 kHz audio. KittenML calls it the world's first ternary text-to-speech model, a claim we have not checked. The linear weights are ternary, so each value in a group of 128 is one of minus scale, zero or plus scale, and the default file packs them into 0.95 GB while the card says it produces bit-identical audio to the 3.47 GB bf16 file. A third file, emb4, shrinks the token embedding to 4 bits for about 0.49 GB and is the only lossy packing of the language model.
It ships 115 built-in voices and clones a voice from 5 to 30 seconds of a single speaker without fine-tuning. The card calls expression control beta. A leading tag sets one of ten emotions for a line, inline tags add ten vocal events such as laughs, sighs and gasps, and triple parentheses stress a word, and only those 20 tags are recognized. Language coverage is stated three ways, with named voices in 16 languages on the model card, 20 languages in the GitHub README and 10 on the homepage, so test the language you need, and pass the card's normalize=False option for non-English text because the text normalizer is tuned for English. The README asks for a CUDA GPU with roughly 8 GB free or about 6 GB of RAM on a CPU. It runs on CPU out of the box, with GGUF weights for a llama.cpp fork as the fastest route, an optional MLX backend for Apple silicon, a vLLM backend and a smaller 39 MB distilled decoder that trades fidelity for footprint.
The weights use the Stellon Labs Community License, which is free for research, non-commercial and limited commercial use and ends the commercial grant once you or your affiliates pass US$1 million in annual revenue or total funding, while the Python package is Apache 2.0. KittenML also offers a hosted API. We found no independent quality comparison, so the README's "SOTA" wording is the developer's own. In the 30 days to 11 November 2026, Hugging Face counted about 66,500 downloads.
Good for
- Voice features on a laptop or a CPU-only server, since the default weights are under 1 GB
- Voice agents, narration and characters that need emotion tags plus laughs, sighs and pauses
- Cloning a voice from a 5 to 30 second clip without fine-tuning
- Startups and small teams under US$1 million in revenue and funding that want free commercial use
Watch out for
- Commercial use needs registration with Stellon Labs, and the grant ends at US$1 million in annual revenue or total funding
- The license bars using the model or its outputs to train or distill other models, except fine-tunes and LoRAs, and requires a "Powered by KittenML" line when you distribute
- Language support is unclear, with 10, 16 and 20 languages stated in different places, and non-English text needs normalize=False
- Emotion control is in beta and steers delivery rather than guaranteeing it
- Voice cloning can imitate real people, so get consent first
Performance claims come from the developer and have not been independently verified.
How much memory does it need?
Memory figures for each variant of KittenTTS 2. Image, video, speech and music pipelines also load a text encoder and a decoder, so the memory needed can be larger than the main checkpoint.
| Variants | Parameters | Download size | Memory needed | Fits on |
|---|---|---|---|---|
| Packed ternary (default) TERNARY | 1.7B | 0.95 GB | 6 GB | Laptop or phone |
| emb4 (4-bit embedding, lossy) TERNARY | 1.7B | 0.49 GB | 6 GB | Laptop or phone |
| Full bf16 BF16 | 1.7B | 3.47 GB | 6 GB | Laptop or phone |
Download size covers every component needed to run the model, such as the generator, text encoder and decoder. Memory needed is a stated or published measured minimum where one exists; otherwise it is the download size plus about 15%. Larger images, longer clips and higher resolutions need more. The developer states a minimum of 6 GB of GPU memory.
The license in plain words
The license for the KittenTTS 2 weights from KittenML, whose text is dated 5 July 2024. Research, non-commercial and limited commercial use are free. Commercial users must register with Stellon Labs, and the commercial grant ends once you or your affiliates pass US$1 million in annual revenue or in total funding.
- Free commercial use only below a revenue cap
- Attribution or model name must be shown
- Acceptable-use policy applies
More about Stellon Labs Community
See also
Frequently asked questions
- What is KittenTTS 2?
- KittenTTS 2 is a text-to-speech model small enough to run on an ordinary computer's processor. It has 115 built-in voices, can copy a voice from a short recording, and lets you add emotions and sounds such as laughs with simple tags. It is free for research and personal use, and free for businesses until they pass US$1 million in revenue or funding.
- Can I use KittenTTS 2 commercially?
- Yes, with conditions. KittenTTS 2 is released under the Stellon Labs Community, which allows commercial use but adds extra rules: Free commercial use only below a revenue cap; Attribution or model name must be shown; Acceptable-use policy applies. Read the license text before you ship a product.
- Is KittenTTS 2 open source?
- KittenTTS 2 is open-weight, not open source in the strict sense. You can download the model files, but the Stellon Labs Community adds conditions that the Open Source Initiative definition does not allow.
- Can I run KittenTTS 2 on my own computer?
- Yes. The smallest variant, Packed ternary (default), needs about 6 GB of memory at 4-bit precision, which fits a modern laptop or a mid-range graphics card.
- When was KittenTTS 2 released?
- KittenTTS 2 was released by KittenML (Stellon Labs) on October 2026.
Official links
Sources
Specifications come from these primary sources. Benchmark claims, where mentioned, are reported by the developer.
- KittenTTS 2 model card (Hugging Face) huggingface.co
- KittenTTS repository: features, requirements and CPU, MLX and vLLM usage (GitHub) github.com
- Stellon Labs Community License text (Hugging Face) huggingface.co
- KittenML homepage kittenml.com
Release date and facts last verified November 11, 2026.
Related models
| Model | Released | Parameters | Context | License | Smallest variant needs |
|---|---|---|---|---|---|
| Pocket TTS Kyutai | 18M – 100M | – | CC BY 4.0 (Permissive) | 0.127 GB Laptop or phone | |
| Supertonic 3 Supertone | 99M | – | OpenRAIL-M (Conditional) | 0.46 GB Laptop or phone | |
| Kokoro hexgrad | 82M | – | Apache 2.0 (Permissive) | 0.38 GB Laptop or phone | |
| VoxCPM2 OpenBMB | 2.29B | – | Apache 2.0 (Permissive) | 8 GB Laptop or phone |