MoshiVis v0.1 - a kyutai Collection

kyutai 's Collections

CASA

Moshi v0.1 Release

MoshiVis v0.1

updated 3 days ago

MoshiVis is a Vision Speech Model built as a perceptually-augmented version of Moshi v0.1 for conversing about image inputs

Vision-Speech Models: Teaching Speech Models to Converse about Images

Paper • 2503.15633 • Published Mar 19 • 2

Note MoshiVis preprint on arxiv
kyutai/Babillage

Viewer • Updated Mar 21 • 465k • 151 • 12

Note Evaluation Benchmark for Vision Speech Models, based on COCO-Captions, OCR-VQA and VQAv2
kyutai/moshika-vis-pytorch-bf16

Updated Jun 18 • 56

Note `bfloat16` weights for the Pytorch backend
kyutai/moshika-vis-candle-bf16

Updated Mar 18 • 1

Note `bfloat16` weights for the rust backend
kyutai/moshika-vis-candle-q8

8B • Updated Mar 18 • 12 • 1

Note 8-bits quantised weights for the rust backend
kyutai/moshika-vis-mlx

Updated Mar 18 • 2

Note weights for the MLX backend (`bfloat16`, 8-bits and 4-bits quantised)