Kyutai

weights

open(PyTorch, MLX and Candle checkpoints, ungated)

data

described(7M hours of audio, Fisher and a Kyutai-collected set

code

partial(inference, server and a separate fine-tuning repository)

license

CC-BY-4.0(Moshi v0.1 release weights)

research-checkpoint

moshika-rl-seamless(2026, gated CC-BY-NC-4.0 fine-tune published with the Interactivity Alignment paper

Moshi's current release, the v0.1 Moshiko and Moshika checkpoints for PyTorch, MLX and Candle, is CC-BY-4.0, an attribution-only license, with code under MIT and Apache-2.0. Kyutai describes its 7 million hours of training audio without releasing them and publishes inference code and a separate fine-tuning repository rather than the training pipeline. The gated, non-commercial moshika-rl-seamless checkpoint belongs to Kyutai's separate Interactivity Alignment research release and does not set the score.

• https://huggingface.co/api/collections/kyutai/interactivity-alignment-6a1fe5f14879ce66956362e2 recorded 2026-09-27

  Collection "Interactivity Alignment", "Full-duplex speech models post-trained with reinforcement learning for improved conversational interactivity.": paper 2606.11167, kyutai/moshika-rl-seamless, kyutai/personaplex-rl-seamless and a samples dataset.

• https://huggingface.co/api/collections/kyutai/moshi-v01-release-66eaeaf3302bef6bd9ad7acd recorded 2026-09-27

  Collection "Moshi v0.1 Release", "MLX, Candle & PyTorch model checkpoints released as part of the Moshi release from Kyutai": paper 2410.00037, kyutai/mimi and the moshiko and moshika pytorch, mlx and candle checkpoints; moshika-rl-seamless is not among them.

• https://huggingface.co/api/models/kyutai/moshika-rl-seamless recorded 2026-09-27

  kyutai/moshika-rl-seamless: "license:cc-by-nc-4.0", "gated":"auto", "base_model:finetune:kyutai/moshika-pytorch-bf16", "arxiv:2606.11167", "createdAt":"2026-06-02T08:37:38.000Z".

• https://huggingface.co/api/models/kyutai/moshiko-pytorch-bf16 recorded 2026-09-27

  "license:cc-by-4.0" tag, "gated":false, with model.safetensors in the file list.

• https://huggingface.co/kyutai/moshiko-pytorch-bf16/raw/main/README.md recorded 2026-09-27

  Training Data section: "Unsupervised audio dataset:** used for pre-training, this is a collection of 7 million hours of readily available audio content"; "A dataset of 170 hours of natural and scripted conversation between multiple pairs of participants, collected by Kyutai."

• https://raw.githubusercontent.com/kyutai-labs/moshi/HEAD/README.md recorded 2026-09-27

  "The weights for the models are released under the CC-BY 4.0 license."; "If you want to fine tune Moshi, head out to [kyutai-labs/moshi-finetune]".

Hugging Face downloads of the three most-used checkpoints. The moshi package is marked as not a separate channel because it loads these repositories.

• https://huggingface.co/api/models?author=kyutai&search=moshi recorded 2026-09-27

  Trailing-30-day downloads: moshiko-pytorch-bf16 179,111, moshiko-candle-q8 22,422, moshika-pytorch-bf16 4,342.

Moshi holds a live two-way spoken conversation, which no recognizer or synthesizer here does, and it was the first open model to do so in real time.

• https://huggingface.co/kyutai/moshiko-pytorch-bf16/raw/main/README.md recorded 2026-09-27

  "Moshi is the first real-time full-duplex spoken large language model".

• https://raw.githubusercontent.com/kyutai-labs/moshi/HEAD/README.md recorded 2026-09-27

  "is a speech-text foundation model and **full-duplex** spoken dialogue framework"; "with a practical overall latency as low as 200ms on an L4 GPU."