mehdi-hf · Released, not on our provider

Cilla Mini (2026): what the Persian voice-cloning text-to-speech model is, its license, speed and how to run it

Cilla Mini is an 18.2M-parameter Persian text-to-speech model by mehdi-hf on Hugging Face, built on Kyutai’s Pocket TTS, that clones a voice from a 3 to 10 second audio prompt, writes 24 kHz audio, and is gated and licensed CC BY-NC 4.0 for non-commercial use only.

Released, not on our providerLast updated:

Use it now

Chat

This chat starts with Qwen3.8 Flash, not Cilla 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
mehdi-hf
Type
Text
Released
2026-11-16
OpenRouter id
None

Checked

Cilla Mini: facts, with sources

  • Hugging Face’s record of the mehdi-hf/cilla-mini repository shows it was created on November 16, 2026 (UTC) and last modified the same day; its sibling mehdi-hf/cilla-lite was created on November 14, 2026. The model card gives no release date, so we say "repository created" rather than "released".Source: huggingface.co · checked 2026-11-19
  • The model card is titled "Cilla Mini: Persian text to speech (18.2M)". It describes the 18M model for CPUs, phones and browsers: Cilla shrunk with Kyutai’s small-model recipe (8 layers, a 64-dim interface, a one-step drifting sampler head and the 2.6M Mimi decoder), with guidance baked in. It speaks Persian in the voice of a short audio prompt and is built on Pocket TTS by Kyutai. Output is 24 kHz mono; the language tag is Persian (fa).Source: huggingface.co · checked 2026-11-19
  • Parameter counts as the card gives them: 18.2M are used for generation (a 15.6M flow language model with 8 layers and the sampler head, plus the 2.6M Mimi decoder), while the weights file holds 27.9M because it also carries the 9.7M Mimi encoder, which runs once per voice prompt. Hugging Face’s own count is 27,925,618 float32 parameters. The main weights file is model.safetensors, 107 MB, in the Pocket TTS format.Source: huggingface.co · checked 2026-11-19
  • License: the card says "Gated, non-commercial (CC BY-NC 4.0)" and, in its licence section, "CC BY-NC 4.0: non-commercial use only. Commercial use needs a separate licence", to be requested through a contact address printed on the card. Hugging Face’s metadata also lists license cc-by-nc-4.0. The Mimi codec and the 2.6M decoder it uses are Kyutai’s, and the card credits Kyutai’s Pocket TTS as the base.Source: huggingface.co · checked 2026-11-19
  • Access: Hugging Face marks the repository as gated with automatic approval. Anyone can read the model card, but to see or download the files you must log in and agree to share your Hugging Face information and accept the conditions. The card’s own commands start with hf auth login and say to ask for access on the page first. This is why we could only read the rendered card, not the raw README or the phonon folder’s README.Source: huggingface.co · checked 2026-11-19
  • Voices: the card says the voice is cloned from a prompt, a whole short clip of 3 to 10 seconds that ends at a pause, in a quiet room; prompts cut mid-word make the model sound broken. The repository ships three prompts: parisa.wav (the voice of Parisa, "recorded by her and used here with her agreement"), voice1.wav (a woman) and voice2.wav (a man). In the maker’s check every clip made from each of them was closest to that prompt.Source: huggingface.co · checked 2026-11-19
  • On cloning, the card is plain: voice cloning is not exact (it follows the prompt’s timbre and style, not every detail); speech made with it should be labelled as synthetic; and do not clone a voice without the person’s consent. The card’s numbers measure intelligibility and a coarse speaker match, and it says they do not show whether a particular voice is reproduced.Source: huggingface.co · checked 2026-11-19
  • Limits listed on the card: Persian only; ezafe mistakes happen, mostly on rare words and in long noun phrases; only Persian script is spoken, and Latin-script words and other foreign characters such as emoji are dropped without a warning, so names must be written in Persian letters; digits are read as Persian words; long text must be spoken sentence by sentence or words are repeated or lost; the Phonon files hold three voices and that runtime’s own voice cloning cannot be used with this model.Source: huggingface.co · checked 2026-11-19
  • The card’s intended use is speech on the user’s own device: offline and private assistants, screen readers and other accessibility tools, games, and apps that cannot wait for a cloud round trip; run on a Mac or in a browser tab, it says nothing leaves the device. The maker’s demo Space runs the model inside the tab with WebAssembly, and the card says the text is not sent anywhere and the page asks before it downloads the 107 MB.Source: huggingface.co · checked 2026-11-19
  • Cilla Mini is not on OpenRouter’s model list (checked November 19, 2026), so this site cannot run it, and this site’s chat is text in, text out: it does not generate speech or audio. For text chat with small models use Qwen3.8 Flash or Inkling 2 Small instead.Source: openrouter.ai · checked 2026-11-19 · second-hand

Our table

Cilla Mini speed: the maker’s figures on a cloud CPU, an Apple M1 Pro and a browser

  • Speed, as the maker measured it ("real time" is 1x): 3.79x for the Python pocket-tts package on one thread of a cloud x86 CPU (Intel Xeon at 2.80 GHz); about 12 to 23x for the Phonon runtime (ptts) on one core of an Apple M1 Pro, with first audio in 30 to 95 ms; and about 9 to 17x in a Chrome tab on an M1 Pro through WebAssembly, short texts being at the slow end. These are the maker’s figures and we did not run them.Source: huggingface.co · checked 2026-11-19

Speed table

How fast Cilla Mini runs where, as printed on its model card

Cilla Mini is built to be small enough for a CPU, a phone or a browser tab, and the card backs that with three speed numbers, where 1x means audio is made as fast as it plays. Even the slowest, one thread of a cloud CPU through the Python package, is 3.79x. The Mac runtime and the browser are several times faster. If you want a Persian voice for an offline app, that is the point of the model; if you want text answers in Persian from a model, use Qwen3.8 Flash or Inkling 2 Small here. An English voice model of similar spirit is tiny voca.

Our table: audio made per second of work, as printed on the Cilla Mini Hugging Face card and read on November 19, 2026; they are the maker’s own figures and we did not run them · checked 2026-11-19
Where it runsAudio per second of workNote
Python (pocket-tts), one thread of a cloud x86 CPU (Intel Xeon, 2.80 GHz)3.79xReleased pocket-tts 4.0.1 or newer
Phonon runtime (ptts), one core of an Apple M1 ProAbout 12 to 23xFirst audio in 30 to 95 ms
Browser tab (WebAssembly), Chrome on an M1 ProAbout 9 to 17xShort texts are at the slow end

Source: huggingface.co

Our table

Cilla Mini quality: word error, character error and speaker similarity by the maker’s check

  • The maker’s evaluation: a Persian speech recognizer transcribed real recordings and voices cloned from other clips of the same speaker (step 100,000, EMA, guidance built in). The recognizer scores 8.6% word error on the real recordings, which the card calls the floor. For the model: word error 10.4%, character error 3.9% and speaker similarity 0.953. The card says these show intelligibility and a coarse speaker match, not whether a particular voice is reproduced or how often the ezafe is right, which needs listening.Source: huggingface.co · checked 2026-11-19

Evaluation table

Cilla Mini evaluation numbers from its model card

The card’s check is honest about its limits. The recognizer itself gets 8.6% of words wrong on real recordings, so the 10.4% for the cloned voices should be read next to that floor, not next to zero. Speaker similarity of 0.953 says the output follows the voice in the prompt, which fits the card’s statement that cloning is not exact. None of this measures how natural the speech sounds or how often the Persian ezafe is right; the card says that needs listening. For the other voice model on this site, see tiny voca, whose card prints its own English checks.

Our table: figures as printed on the Cilla Mini Hugging Face card, read on November 19, 2026; a Persian speech recognizer transcribed real recordings and voices cloned from other clips of the same speaker; they are the maker’s own figures and we did not run them · checked 2026-11-19
MeasureCilla Mini (cloned voices)
Real recordings, word error rate (the recognizer’s floor)8.6% (recognizer alone on real recordings)
Word error rate (WER)10.4%
Character error rate (CER)3.9%
Speaker similarity0.953

Source: huggingface.co

Our table

How to run Cilla Mini: license, access, Python, Mac and browser

  • Ways to run it, from the card: the Python route (pip install "pocket-tts>=4.0.1", hf auth login, hf download of the repository without the large weights, then speak_fa.py with a prompt, a text and an output file, which fetches the weights once); the Phonon route for a Mac (a Rust runtime with no PyTorch and no GPU, set up with the folder’s setup.sh); and the browser demo Space mehdi-hf/cilla-mini-demo. The card says a sentence must be spoken one at a time from the same prompt and not cut into 50-token pieces, which made the models repeat words.Source: huggingface.co · checked 2026-11-19
  • Hugging Face lists the demo Space mehdi-hf/cilla-mini-demo as a static Space with the short description "Persian text to speech that runs in your browser", license cc-by-nc-4.0, containing the model file, the three voice files and the web code.Source: huggingface.co · checked 2026-11-19
  • The sibling Cilla Lite is described on its own card as the small 6-layer model for CPUs, distilled from Cilla, with about 100M parameters used for generation (the file holds 110M), 24 kHz mono, Persian, gated and CC BY-NC 4.0, and it names a larger Cilla (326M) as its sibling. Its speed is given as 1.88x real time on one thread of a cloud x86 CPU, against 3.79x for Cilla Mini.Source: huggingface.co · checked 2026-11-19

Ways to run it

The ways the maker documents to run Cilla Mini and what each needs

Everything starts with access: the repository is gated, so you log in to Hugging Face, accept the conditions and then download. The licence is non-commercial, so a product that earns money needs the maker’s separate licence first. The browser demo is the quickest way to hear it, because the card says it runs inside the tab and sends no text anywhere. This site cannot run Cilla Mini; for text models that can, compare Inkling 2 Small and Qwen3.8 Flash.

Our table: from the Cilla Mini Hugging Face card and the demo Space listing, read on November 19, 2026; we did not run any of these commands, and the gated files and the Phonon README could not be read · checked 2026-11-19
RouteHowWhat it needs
Python (pocket-tts)Log in, download the code, prompts and config, then run speak_fa.py with a prompt clip, a Persian text and an output file; it fetches the weights oncepocket-tts 4.0.1 or newer; Hugging Face access; keep the three flags in config.yaml set to false
Phonon runtime (Mac)The phonon folder holds the model laid out for a Rust runtime with no PyTorch and no GPU; setup.sh downloads the runtime and checks checksums; say_fa.py and run_web.sh give a command line and a web pageApple silicon (the card’s timings are on an M1 Pro); three built-in voices; the runtime’s own cloning cannot be used with this model
Browser demoThe Space mehdi-hf/cilla-mini-demo runs the model in the tab with WebAssembly; text is not sent anywhere; it asks before downloading 107 MBA browser; open it in its own tab for the multi-threaded mode
Commercial useNot allowed under CC BY-NC 4.0; needs a separate licence from the makerContact address printed on the card

Source: huggingface.coSource: huggingface.co

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 Cilla 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

FAQ

Cilla Mini FAQ

What is Cilla Mini?

Cilla Mini is a Persian text-to-speech model: you give it Persian text and a short voice prompt and it speaks the text in that voice. The card calls it the 18M model for CPUs, phones and browsers, made by shrinking Cilla with Kyutai’s small-model recipe, and built on Pocket TTS by Kyutai. About 18.2M parameters are used to generate speech; the file holds 27.9M. It is not a chat model.

Can Cilla Mini clone voices?

Yes, from a prompt. The card says the voice is cloned from a whole short clip of 3 to 10 seconds that ends at a pause, in a quiet room, and that cloning is not exact: it follows timbre and style, not every detail. Cloning a real person’s voice without consent can mislead listeners and harm that person. The card asks you to label the speech as synthetic and not to clone a voice without the person’s consent; its Parisa sample is used with her agreement.

Is Cilla Mini non-commercial, and what is the exact licence?

Yes. The card says "Gated, non-commercial (CC BY-NC 4.0)" and that CC BY-NC 4.0 means non-commercial use only; commercial use needs a separate licence, requested through a contact address printed on the card. Hugging Face’s metadata lists cc-by-nc-4.0. The Mimi codec and decoder it uses are Kyutai’s. We are not lawyers; read the card before any paid use.

Why is Cilla Mini gated and how do I get access?

Hugging Face marks it as gated with automatic approval: anyone can read the card, but to see or download the files you must log in, agree to share your Hugging Face information and accept the conditions. The card’s commands start with hf auth login. We could read only the rendered card, so details that live in the files, such as the Phonon README, are not covered here.

How fast is Cilla Mini and what hardware does it need?

By the maker’s own figures, which we did not test, it runs at 3.79x real time on one thread of a cloud x86 CPU through the Python package, about 12 to 23x on one core of an Apple M1 Pro with the Phonon runtime, and about 9 to 17x in a Chrome tab. No GPU is needed. The weights file is 107 MB and the repository needs pocket-tts 4.0.1 or newer for the Python route.

How good is Cilla Mini?

On real Persian recordings, with voices cloned from other clips of the same speaker, the card reports 10.4% word error, 3.9% character error and 0.953 speaker similarity; the recognizer alone gets 8.6% word error, which the card calls the floor. The card says this does not show how natural it sounds or how often the ezafe is right, and lists ezafe mistakes, dropped Latin-script words and long-text problems as limits.

Can I use Cilla Mini on this site?

No. It is not on OpenRouter’s model list and this site does not generate speech; its chat answers in text. You can chat with a small model instead: on this site a typical reply to Inkling 2 Small costs about 1 credit (a credit is about $0.01 in the $5 pack), billed by its real token cost, at least 1 credit. Qwen3.8 Flash is another, and tiny voca is an English voice model.

Is this the official Cilla Mini website?

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