KittenML · Released, not on our provider

Kitten TTS 2 Mini (2026): the open 82M-parameter voice model, its languages, voices, CPU needs and how it differs from Pico

Kitten TTS 2 Mini is KittenML’s open text-to-speech model with 82M parameters, 16 languages, 118 voices and 24 kHz output, published as Apache-2.0 weights on Hugging Face that run on a CPU, and this site cannot run it because it is a text chat, not a speech generator.

Released, not on our providerLast updated:

Use it now

Chat

This chat starts with Qwen3.8 Flash, not Kitten TTS 2 Mini.

≈ 1 credit per reply · See all models

New accounts get 5 free credits once, valid for 7 days. They work on the lower-priced models here; the priciest ones need credits you bought.

Start chatting — 5 free credits
Maker
KittenML
Type
Text
Released
No date given
OpenRouter id
None

Checked

Kitten TTS 2 Mini: facts, with sources

  • The Hugging Face repository KittenML/kitten-tts-2-mini was created on November 13, 2026 (its initial commit and the "Kitten TTS 2 Mini" commit are both from that day, UTC), and the int4 ONNX graphs were added on November 18, 2026. The model card gives no release date or announcement, so we treat November 13 as the day the repository was created, not as a launch date.Source: huggingface.co · checked 2026-11-19
  • The model card lists 82M parameters, 16 languages (English, German, Spanish, Italian, Portuguese, Russian, Hindi, Danish, Finnish, French, Japanese, Korean, Dutch, Swedish, Swahili and Turkish), 118 voices and 24 kHz output. The repository ships ONNX graphs with int8 weights (82.7 MB) or int4 weights (47.5 MB), both of which run on a CPU without PyTorch, plus float32 PyTorch weights (the file listing shows about 327 MB).Source: huggingface.co · checked 2026-11-19
  • Kitten’s own docs page for the Nano, Pico and Mini family says Mini is a single model that covers all 16 languages (Nano and Pico ship eight models each, one per language). English text goes through text normalisation, a lexicon, CMUdict and a 0.75M-parameter BART grapheme-to-phoneme model; other languages use espeak-ng, and Japanese needs the extra install kittenml[ja]. Long text is split at sentence ends into chunks of at most 350 phonemes, joined with 0.15 seconds of silence.Source: github.com · checked 2026-11-19
  • To run it, install the kittenml Python package and load KittenTTS("KittenML/kitten-tts-2-mini"); generate() takes the text, an optional voice name, a language code and a speed (1.0 by default, larger is faster) and returns audio you can write at 24,000 Hz. weights="int4" selects the smaller graphs (needs onnxruntime 1.20 or newer), backend="torch" runs the float32 weights, and device="cuda" with that backend uses an NVIDIA GPU. The docs also list a streaming call that returns audio chunk by chunk. Neither the card nor the docs give a memory, latency or real-time-factor figure for Mini.Source: github.com · checked 2026-11-19
  • The Mini card does not mention voice cloning, reference audio or expression tags: you choose one of the built-in voices by name or by its description.Source: huggingface.co · checked 2026-11-19
  • The card says the models, voice packs and bundled grapheme-to-phoneme files are under Apache-2.0, and it states no commercial-use limit beyond that license. The GitHub README adds that the code is Apache-2.0 too but that each model repository carries its own terms, so the license of the exact repository you download is the one that counts.Source: huggingface.co · checked 2026-11-19
  • The larger flagship KittenTTS 2 (1.7B parameters, in-context voice cloning from 5-30 seconds of a single speaker, 115 built-in voices) is a different model with different terms: its card names the Stellon Labs Community License, under which research, non-commercial and limited commercial use are free of charge and the commercial grant ends once you or your affiliates pass USD $1,000,000 in annual revenue or total cumulative funding. That does not apply to the Apache-2.0 Mini card, but it is why a search for "Kitten TTS 2" can lead to either license.Source: huggingface.co · checked 2026-11-19
  • The GitHub README calls Kitten-TTS "a series of sota open-source text-to-speech models, built by KittenML" and lists Nano, Pico and Mini as small multilingual models shipped as int8 or int4 ONNX for CPU. The README’s claim of real-time CPU speed is made for the 1.7B flagship, not for Mini, and we did not test either.Source: github.com · checked 2026-11-19
  • Kitten TTS 2 Mini is not on OpenRouter’s model list (checked November 19, 2026), and this site is a text chat with language models, so it cannot generate speech with it. This site can chat with small text models such as Qwen3.8 Flash, Inkling 2 Small and Hy4 Flash.Source: openrouter.ai · checked 2026-11-19 · second-hand

Our table

Kitten TTS 2 Mini vs Pico vs Nano: parameters, languages, voices and file size as printed

  • The Pico card lists 7.2M parameters per language model, 8 languages (English, German, Spanish, Italian, Portuguese, Russian, Hindi and French), 30 voices and 24 kHz output, with int8 weights of 8.0 MB or int4 weights of 4.0 MB per language; it is under Apache-2.0 and, like Mini, runs on a CPU through ONNX Runtime.Source: huggingface.co · checked 2026-11-19
  • The Nano card lists 17M parameters per language model, the same 8 languages as Pico, 37 voices and 24 kHz output, with int8 weights of 17.4 MB or int4 weights of 8.8 MB per language, under Apache-2.0.Source: huggingface.co · checked 2026-11-19

Mini vs Pico

Kitten TTS 2 Mini, Pico and Nano side by side

People who search for Kitten TTS 2 Mini usually want to know how it differs from the other small Kitten models. Mini is the largest of the three and the only one that covers 16 languages in a single model; Pico is the smallest, at 7.2M parameters per language and 4.0 MB per language in int4. The cards print the same sample rate, license and install route for all three, and only Pico and Nano have a printed quality table, so we compare size and coverage here, not sound quality. Other open speech models we cover are ZONOS3-Mini, which clones voices, and XTTS v2, which has a non-commercial license. Microsoft’s Aion 1.0 is a small on-device model of a different kind: it writes text, not speech.

Our table: figures from the three Hugging Face model cards, the Kitten docs page and the Hugging Face repository list, read on November 19, 2026; sizes are per model file, and Pico and Nano have one file per language · checked 2026-11-19
ItemKitten TTS 2 MiniKitten TTS 2 PicoKitten TTS 2 Nano
Parameters82M7.2M per language model17M per language model
Languages16 (en, de, es, it, pt, ru, hi, da, fi, fr, ja, ko, nl, sv, sw, tr)8 (en, de, es, it, pt, ru, hi, fr)8 (en, de, es, it, pt, ru, hi, fr)
Models in the repository1, for all languages8, one per language8, one per language
Voices (card headline)1183037
English voices11410
int8 ONNX size82.7 MB8.0 MB per language17.4 MB per language
int4 ONNX size47.5 MB4.0 MB per language8.8 MB per language
Sample rate and license24 kHz, Apache-2.024 kHz, Apache-2.024 kHz, Apache-2.0
Printed quality tableNone foundint4 error-rate table in the docsint4 error-rate table in the docs
Repository createdNovember 13, 2026November 13, 2026November 13, 2026

Source: huggingface.coSource: huggingface.coSource: huggingface.coSource: github.comSource: huggingface.co

Our table

Kitten TTS 2 Mini languages and voices: how many voices each language has

  • The card lists voices as first names with a short description: 11 English voices (such as Eleanor, "Somber historian", and Max, "Energetic sports announcer") and, for the other languages, plain labels such as "Female 1" and "Male 2". Nine of the English-named voices also appear in the German, Spanish, Italian, Portuguese and Russian lists; the card says they speak those languages under the same name when you pass language=, and English without it. Adding up the card’s tables gives 118 entries, the same as its headline number, with those English-named voices counted once per list they appear in.Source: huggingface.co · checked 2026-11-19

Voices by language

Voices listed on the Kitten TTS 2 Mini card, by language

The card gives every voice a first name; the voices for German, Spanish, Italian, Portuguese and Russian have four own names each, followed by nine of the English-named voices, which is why those five lists have 13 entries. Voice descriptions such as "Somber historian" are labels for fictional characters on the card, not real people. The model only speaks; for a model that writes the text first, try Qwen3.8 Flash or Hy4 Flash in the chat.

Our table: counted by us from the voice tables on the Kitten TTS 2 Mini card, read on November 19, 2026; the five languages with 13 entries include nine English-named voices each · checked 2026-11-19
LanguageEntries on the cardVoice names for this language
English (en)11Eleanor, Martha, Serena, Iris, Grace, Edith, Diana, Otis, Vincent, Max, Danny
German (de)13Lena, Greta, Jonas, Felix, plus nine English-named
Spanish (es)13Lucia, Carmen, Javier, Diego, plus nine English-named
Italian (it)13Giulia, Chiara, Marco, Luca, plus nine English-named
Portuguese, Brazil (pt)13Beatriz, Camila, Rafael, Thiago, plus nine English-named
Russian (ru)13Anastasia, Natasha, Dmitri, Ivan, plus nine English-named
Hindi (hi)4Priya, Ananya, Arjun, Vikram
Danish (da)3Freja, Mads, Rasmus
Finnish (fi)5Aino, Helmi, Eero, Juho, Ville
French (fr)8Camille, Chloe, Manon, Louis, Antoine, Julien, Mathis, Etienne
Japanese (ja)5Yuki, Sakura, Haruka, Aoi, Haruto
Korean (ko)3Seoyeon, Minjun, Jihoon
Dutch (nl)5Sanne, Fenna, Daan, Bram, Joris
Swedish (sv)3Astrid, Elsa, Oskar
Swahili (sw)4Zawadi, Neema, Baraka, Juma
Turkish (tr)2Elif, Zeynep

Source: huggingface.co

Our table

Kitten TTS 2 Mini has no quality table: Pico and Nano int4 error rates are the only numbers Kitten prints

  • Kitten’s docs report an int4 check on 60 held-out sentences per language, measured as Whisper large-v3 character error rate with the float32 value in brackets, for Pico and Nano only: for example English 1.17 (0.86) for Pico and 1.08 (1.08) for Nano, Hindi 10.72 (10.70) and 13.21 (13.13). They add that speaker-embedding similarity to float32 is 0.996 to 0.999 and that 95 to 99% of phoneme durations are identical to float32’s. We found no such table for Mini.Source: github.com · checked 2026-11-19

Maker’s figures

Kitten TTS 2 Mini’s siblings: character error rate of int4 Pico and Nano, float32 in brackets

This is the one place where Kitten prints how a small model sounds to a recogniser, and Mini is missing from it. Read it as a check that int4 compression barely changed Pico and Nano against their float32 versions, not as a ranking of the three models: the sentences and the languages are the same, but the models are different sizes. Hindi stands out as the hardest language for both models. If you want text answers instead of audio, Inkling 2 Small is a small model you can chat with on this site.

Our table: values as printed on the Kitten docs page for the small models, read on November 19, 2026; Whisper large-v3 character error rate on 60 held-out sentences per language, lower is better; the docs give no table for Mini and we did not run any of it · checked 2026-11-19
LanguagePico int4 (float32)Nano int4 (float32)
English1.17 (0.86)1.08 (1.08)
German1.64 (1.78)1.28 (1.63)
Spanish0.98 (0.91)0.89 (0.83)
Italian0.94 (0.66)0.69 (0.61)
Portuguese4.08 (4.10)3.93 (3.87)
Russian5.79 (5.87)6.48 (6.50)
Hindi10.72 (10.70)13.21 (13.13)
French2.13 (2.05)2.13 (1.94)

Source: github.com

Get told when it is available

Sign in and press the button: we store your account and this model, nothing else. No emails are sent yet.

Nearest available

Alternatives to Kitten TTS 2 Mini you can use now

Qwen3.8 Flash

Alibaba

$0.15 / $0.47 (per million tokens)

  • Alibaba’s price for qwen3.8-flash is $0.15 / $0.47 per million input / output tokens (Singapore, EU, Hong Kong and US regions).Source: docs.qwencloud.com · checked 2026-11-13
  • Listed on OpenRouter as qwen/qwen3.8-flash at $0.15 / $0.47 per million tokens.Source: openrouter.ai · checked 2026-11-13

Inkling 2 Small

Thinking Machines Lab

$0.30 / $1 (per million tokens)

  • Thinking Machines Lab announced Inkling 2 Small on November 5, 2026, and made it available on its Tinker platform.Source: thinkingmachines.ai · checked 2026-11-14
  • A mixture-of-experts model with 594B total parameters and 17B active, with image input and variable effort.Source: thinkingmachines.ai · checked 2026-11-14
  • Text and image input, text output; context window up to 1M tokens; each token is routed to 8 of 384 experts.Source: huggingface.co · checked 2026-11-14

Hy4 Flash

Tencent

$0.04 / $0.14 (per million tokens)

  • Tencent’s Hy4 flash announcement, dated November 16, 2026, presents it as the newest member of the Hy model family. Tencent recommends it for most everyday tasks (deep research, document processing, PowerPoint generation, simple front-end work, computer use) and points complex tasks such as long-horizon software engineering to Hy4. Tencent Cloud’s API guide calls it "a leaner, more efficient everyday workhorse".Source: tencent.com · checked 2026-11-17
  • The Tencent announcement is dated November 16, 2026, and Tencent Cloud’s model, pricing and API pages for hy4-flash were last updated the same day. The announcement says only that Hy4 flash is available now; it gives no separate release date or version name.Source: tencent.com · checked 2026-11-17
  • Tencent says Hy4 flash has 224B total parameters (including 64B Engram parameters), 12B active parameters, a context of up to 1M tokens and multimodal input; the announcement does not state a maximum output or the output types. It says the KV cache is 1/6 the size of Hy4’s and that decoding exceeds 150 tokens per second. These are the maker’s own figures, not tested by us.Source: tencent.com · checked 2026-11-17

FAQ

Kitten TTS 2 Mini FAQ

What is Kitten TTS 2 Mini?

Kitten TTS 2 Mini is a small text-to-speech model from KittenML: it turns text into 24 kHz audio in 16 languages with 118 built-in voices. The Hugging Face card gives 82M parameters and ships int8 (82.7 MB) and int4 (47.5 MB) ONNX files that run on a CPU, plus float32 PyTorch weights. The repository was created on November 13, 2026, and the card gives no launch date.

How is Kitten TTS 2 Mini different from Pico and Nano?

Mini is one 82M-parameter model that covers 16 languages, while Pico (7.2M parameters per language model) and Nano (17M) cover 8 languages with one model per language. Pico is the smallest, at 8.0 MB int8 or 4.0 MB int4 per language, against 82.7 MB and 47.5 MB for Mini. Kitten’s docs print an int4 error-rate table for Pico and Nano but none for Mini, so we cannot rank their sound quality.

What hardware does Kitten TTS 2 Mini need and how do I run it?

The card says both ONNX versions run on a CPU with no PyTorch; the int4 version needs onnxruntime 1.20 or newer. Install the kittenml package, load KittenTTS("KittenML/kitten-tts-2-mini"), call generate with a voice name and write the result at 24,000 Hz. A GPU is optional through the PyTorch backend. The card and docs give no RAM, latency or real-time-factor figure, so we cannot say how fast it is on your machine.

Which languages and voices does it have?

Sixteen languages: English, German, Spanish, Italian, Portuguese (Brazil), Russian, Hindi, Danish, Finnish, French, Japanese, Korean, Dutch, Swedish, Swahili and Turkish. The card lists 118 voice entries, from 2 for Turkish to 13 each for German, Spanish, Italian, Portuguese and Russian, because nine English-named voices appear in those five lists. Japanese needs the extra kittenml[ja] install.

What is the Kitten TTS 2 Mini license, and can I use it commercially?

The Mini card says the models, voice packs and bundled grapheme-to-phoneme files are Apache-2.0 and states no extra commercial-use limit. The larger flagship KittenTTS 2 is different: it uses the Stellon Labs Community License, whose commercial grant ends once you or your affiliates pass USD $1,000,000 in annual revenue or cumulative funding. Read the license file of the exact repository you download.

Does Kitten TTS 2 Mini clone voices?

The Mini card describes no voice cloning: you pick a built-in voice by name. Cloning is a feature of the separate 1.7B flagship KittenTTS 2, which clones from 5-30 seconds of one speaker. If you ever clone a voice, use only a recording of someone who has agreed to it and do not present the result as that person; this note is ours, not KittenML’s.

Can I generate speech with Kitten TTS 2 Mini here?

No. This site is a text chat with language models, so it cannot produce audio with Kitten TTS 2 Mini or any other speech model, and the model is not on OpenRouter’s list. For text, a typical reply to Inkling 2 Small costs about 1 credit on this site (a credit is about $0.01 in the $5 pack), billed by its real token cost, at least 1 credit. Qwen3.8 Flash and Hy4 Flash are other small models.

Is this the official KittenML website?

No. This page is unofficial and not affiliated with or endorsed by KittenML. Facts link to the pages we read.