What is KittenTTS 2?
KittenTTS 2 is a text-to-speech model from KittenML, the brand of Y Combinator-backed Stellon Labs, published on Hugging Face in late September 2026. It is a different kind of model from the 15M to 80M-parameter KittenTTS 0.8 releases: a 1.7-billion-parameter speech language model that writes S3 codec tokens, which a vocoder turns into 24 kHz audio. Its linear weights are ternary (each one of three values), so the default download is 954 MB against 3,469 MB in plain bf16, with the two producing bit-identical audio.
It ships 115 built-in voices, clones a voice from 5 to 30 seconds of one speaker, and takes ten emotion tags, ten vocal-event tags such as <laugh> and triple-parenthesis emphasis, all marked beta. The model card names voices for 16 languages; the GitHub README lists 20. It runs through the kittenml Python package (PyTorch, vLLM on NVIDIA, MLX on Apple silicon) or KittenML's llama.cpp fork for CPUs.
The Python code is Apache 2.0, but the weights are under the Stellon Labs Community License, which ends free commercial use once you pass $1 million in annual revenue or in total funding raised. The GitHub repository behind it has about 16,600 stars.
How much does KittenTTS 2 cost?
| Plan | Price | What you get |
|---|---|---|
| Weights | Free | Stellon Labs Community License; commercial use under $1 million revenue and funding |
| Enterprise licence | Not published | Required above the thresholds |
Weights free under the Stellon Labs Community License for research and for commercial users under $1 million in revenue and in funding, with registration. Hosted API pricing not published.
Is KittenTTS 2 worth it?
KittenTTS 2 is an interesting compression result with too little evidence behind its quality. Packing a 1.7B speech model into ternary weights cuts the download to 954 MB with no loss, and KittenML's MLX notes explain why that matters: generation is one forward pass per 1/25 second of speech and is limited by memory bandwidth, so reading about 1.4 GB per token instead of 3.5 GB is the speed-up. What KittenML does not publish is a word error rate, a listening test or a leaderboard entry, and the model is absent from both Artificial Analysis text-to-speech boards.
The 'runs on a CPU' promise needs reading closely. KittenML's own MLX guide says the default PyTorch backend runs below real time on an Apple-silicon Mac, and Better Stack's write-up found a short sentence taking several seconds there. The fast path is the beta C++ fork: on an eight-thread Xeon it generates 45.7 tokens a second, which at 25 tokens per second of audio is about 1.8 times real time for the language model alone by our arithmetic, before the vocoder runs. On an RTX 3060, vLLM with 4-bit weights cuts a long passage from 45.20 seconds in PyTorch to 10.38.
Cloning is the headline feature, and KittenML's only number is its own: 0.81 speaker similarity when cloning an unseen recording. The weights licence does not mention consent or impersonation; it points to an acceptable-use policy that is only available by contacting Stellon Labs, so the rules on cloning real people are not public.
The licence is the bigger obstacle. Free commercial use requires registering with Stellon Labs and ends once you or your affiliates pass $1 million in annual revenue or in total cumulative funding, which excludes most start-ups after a seed round. Products must display 'Powered by KittenML', outputs may not train foundational generative models, and the grant is revocable. Chatterbox and Kokoro, both on Artificial Analysis's open-weights board, come with far simpler terms.
Where it earns its place
A solo developer building an offline reading app for Linux laptops
Read articles aloud with no cloud calls and no GPU
Build kitten-tts-2-cpp and use the student_w4 decoder, which is 39 MB on disk. Expect a batch CLI rather than a streaming server, and keep inputs chunked: KittenML says the model truncates or drifts into repetition on long passages in one pass.
A university speech lab comparing small cloning models
Measure how much quality ternary packing costs against bf16
Research use is free, and the repository ships the bf16, ternary and 4-bit-embedding packings of the same weights, so one study can compare them directly. KittenML publishes no word error rate, so the lab's own evaluation fills a real gap.
The CTO of a seed-funded start-up with $2 million raised
Check whether KittenTTS 2 can ship in a paid product
Not without a separate licence: total funding above $1 million ends the free commercial grant regardless of revenue. Commercial terms are by request to info@stellonlabs.com, and no prices are published.
Before you commit
- Setup
- pip install kittenml, load KittenML/kitten-tts-2 and call generate with a voice name or a reference clip; no Hugging Face login is needed. KittenML asks for a CUDA GPU with about 8 GB free or a CPU with about 6 GB of RAM, plus about 2 GB of disk.
- Integrations
- PyTorch by default, an optional vLLM backend on NVIDIA (driver 580 or newer), an MLX backend on Apple silicon, a Docker image with an OpenAI-compatible speech API and voice cloning, and the kitten-tts-2-cpp llama.cpp fork with Python bindings.
- Data and privacy
- Stellon Labs Community License: free for research and non-commercial use; commercial use needs registration and stops above $1 million in annual revenue or total funding; 'Powered by KittenML' attribution; no training foundational generative models on outputs; revocable. The bundled pyannote speaker model is MIT and the code is Apache 2.0. Self-hosted, audio never leaves your machine.
- Cost at scale
- Self-hosting is cheap: KittenML's vLLM benchmark uses a single RTX 3060 12 GB, and the C++ runtime holds the transformer weights in about 357 MiB on CPU. The hosted KittenML platform requires sign-in and publishes no prices.
- Limits
- Emotion control is beta and varies by voice. The English text normaliser mangles numbers and dates in other languages unless normalize=False. Streaming is chunk-level, so text under about 380 characters arrives as one chunk; the C++ runtime cannot stream and needs an offline Python step to prepare a cloned voice; its TQ2_1 GGUF does not load in stock llama.cpp.
What's good
- 954 MB lossless ternary download, 469 MB with a lossy embedding
- Voice cloning from 5 to 30 seconds of audio
- 115 built-in voices with emotion and vocal-event tags
- CPU, NVIDIA (vLLM) and Apple-silicon (MLX) backends
What's not
- No published accuracy or listening-test results; absent from Artificial Analysis boards
- Free commercial use ends at $1 million revenue or $1 million funding
- Acceptable-use policy not public
- Default PyTorch backend runs below real time on a Mac
Best for
Hobbyists, researchers and unfunded developers who want offline voice cloning on ordinary hardware and can live with an unproven quality ceiling.
KittenTTS 2 alternatives
KittenTTS 2: common questions
- Is KittenTTS 2 worth it?
- KittenTTS 2 squeezes a cloning speech model into under a gigabyte, but nobody outside KittenML has yet shown how it sounds, and the licence shuts out most funded companies. We rate it 3.4 out of 5.
- How much does KittenTTS 2 cost?
- Weights free under the Stellon Labs Community License for research and for commercial users under $1 million in revenue and in funding, with registration. Hosted API pricing not published.
- What are the downsides of KittenTTS 2?
- No published accuracy or listening-test results; absent from Artificial Analysis boards; free commercial use ends at $1 million revenue or $1 million funding; acceptable-use policy not public; default PyTorch backend runs below real time on a Mac.
- Who is KittenTTS 2 best for?
- Hobbyists, researchers and unfunded developers who want offline voice cloning on ordinary hardware and can live with an unproven quality ceiling.
- How much does KittenTTS 2 cost for a team or in production?
- Self-hosting is cheap: KittenML's vLLM benchmark uses a single RTX 3060 12 GB, and the C++ runtime holds the transformer weights in about 357 MiB on CPU. The hosted KittenML platform requires sign-in and publishes no prices.
- Is KittenTTS 2 safe for company data?
- Stellon Labs Community License: free for research and non-commercial use; commercial use needs registration and stops above $1 million in annual revenue or total funding; 'Powered by KittenML' attribution; no training foundational generative models on outputs; revocable. The bundled pyannote speaker model is MIT and the code is Apache 2.0. Self-hosted, audio never leaves your machine.
- What are the best KittenTTS 2 alternatives?
- The closest alternatives we have reviewed are Kokoro-82M, Chatterbox, Qwen3-TTS and OmniVoice.
Answers checked Oct 28, 2026.