inklight-app/pocket-tts-q8

Pocket TTS, Q8_0

Kyutai's Pocket TTS converted for Inklight, an iPhone app that reads books aloud on the device, and for PocketVoice Reader, a Firefox extension that reads web pages aloud. The FlowLM transformer reads all its weights once for every frame it generates, which makes it memory-bound on a phone; with its linear layers in Q8_0 it reads a quarter as much and renders about 30 % faster, and the file is little more than half the size of the original.

Files

One model.q8.gguf per language, about 129 MB each: english_2026-04, german, french, spanish and portuguese, each the model of the same name from kyutai/pocket-tts-without-voice-cloning (German, French, Spanish and Portuguese as of Kyutai's commit 1e08e6a of 2026-10-01, the versions pocket-tts 4.0.1 uses), with the FlowLM transformer's linear weights in Q8_0 and every other tensor in F16. The Mimi encoder, which turns audio into latents and is only used to clone a voice, is left out; in the release without voice cloning its weights are not usable anyway. Tensor names are those of the original model.safetensors.

Beside the English model, english_2026-04/mimi_small.safetensors (10 MB) holds the Mimi decoder of Kyutai's english_18m_2026-10 model: 2.6 million parameters against the standard decoder's 10.3 million, reading the same latents, since every model in the release shares one Mimi encoder and its normalisation. Inklight uses it in place of the decoder in model.q8.gguf: it costs about a third as much to run, and in a blind comparison on identical latents the owner heard no difference in five of six pairs. It contains only the decoder's tensors (mimi.quantizer.output_proj, mimi.upsample, mimi.decoder_transformer, mimi.decoder) as 32-bit floats, under their original names.

Each file needs the tokenizer and voice embeddings of its language from Kyutai's repository, which are unchanged and not copied here, and the slower voice states in inklight-app/paced-voices load into the English one the same way.

How it was made

With quantize_model in goocy/pocket-tts-ios (branch aloud), a fork of the Rust port UnaMentis/pocket-tts-ios, which also loads the file. It is the same conversion the app used to run on each phone after downloading the original. Against the full-precision model, the transformer's hidden state differs by about 2 % (4.5 % in the worst frame of a 25-second utterance), and in listening tests the differences between renders followed the random seed, not the precision.

Credits and licence

This is an adaptation of Pocket TTS by Kyutai, licensed under the Creative Commons Attribution 4.0 International licence, and is released under the same licence. Kyutai's terms of use apply to it as to the original: it must not be used for voice impersonation or cloning without consent, to deceive, or to produce unlawful or harmful content.

Changes made: the transformer's linear weights were quantized to 8 bits, the other weights converted to 16-bit floats, and the voice-cloning encoder removed. For mimi_small.safetensors: the decoder's tensors were taken from english_18m_2026-10 unchanged, converted to 32-bit floats, and everything else left out. Kyutai does not endorse these files or the apps. They are provided without warranty.