_version_ 1866912067992158208
author Agrawal, Pravesh
Antoniak, Szymon
Hanna, Emma Bou
Bout, Baptiste
Chaplot, Devendra
Chudnovsky, Jessica
Costa, Diogo
De Monicault, Baudouin
Garg, Saurabh
Gervet, Theophile
Ghosh, Soham
Héliou, Amélie
Jacob, Paul
Jiang, Albert Q.
Khandelwal, Kartik
Lacroix, Timothée
Lample, Guillaume
Casas, Diego Las
Lavril, Thibaut
Scao, Teven Le
Lo, Andy
Marshall, William
Martin, Louis
Mensch, Arthur
Muddireddy, Pavankumar
Nemychnikova, Valera
Pellat, Marie
Von Platen, Patrick
Raghuraman, Nikhil
Rozière, Baptiste
Sablayrolles, Alexandre
Saulnier, Lucile
Sauvestre, Romain
Shang, Wendy
Soletskyi, Roman
Stewart, Lawrence
Stock, Pierre
Studnia, Joachim
Subramanian, Sandeep
Vaze, Sagar
Wang, Thomas
Yang, Sophia
author_facet Agrawal, Pravesh
Antoniak, Szymon
Hanna, Emma Bou
Bout, Baptiste
Chaplot, Devendra
Chudnovsky, Jessica
Costa, Diogo
De Monicault, Baudouin
Garg, Saurabh
Gervet, Theophile
Ghosh, Soham
Héliou, Amélie
Jacob, Paul
Jiang, Albert Q.
Khandelwal, Kartik
Lacroix, Timothée
Lample, Guillaume
Casas, Diego Las
Lavril, Thibaut
Scao, Teven Le
Lo, Andy
Marshall, William
Martin, Louis
Mensch, Arthur
Muddireddy, Pavankumar
Nemychnikova, Valera
Pellat, Marie
Von Platen, Patrick
Raghuraman, Nikhil
Rozière, Baptiste
Sablayrolles, Alexandre
Saulnier, Lucile
Sauvestre, Romain
Shang, Wendy
Soletskyi, Roman
Stewart, Lawrence
Stock, Pierre
Studnia, Joachim
Subramanian, Sandeep
Vaze, Sagar
Wang, Thomas
Yang, Sophia
contents We introduce Pixtral-12B, a 12--billion-parameter multimodal language model. Pixtral-12B is trained to understand both natural images and documents, achieving leading performance on various multimodal benchmarks, surpassing a number of larger models. Unlike many open-source models, Pixtral is also a cutting-edge text model for its size, and does not compromise on natural language performance to excel in multimodal tasks. Pixtral uses a new vision encoder trained from scratch, which allows it to ingest images at their natural resolution and aspect ratio. This gives users flexibility on the number of tokens used to process an image. Pixtral is also able to process any number of images in its long context window of 128K tokens. Pixtral 12B substanially outperforms other open models of similar sizes (Llama-3.2 11B \& Qwen-2-VL 7B). It also outperforms much larger open models like Llama-3.2 90B while being 7x smaller. We further contribute an open-source benchmark, MM-MT-Bench, for evaluating vision-language models in practical scenarios, and provide detailed analysis and code for standardized evaluation protocols for multimodal LLMs. Pixtral-12B is released under Apache 2.0 license.
format Preprint
id arxiv_https___arxiv_org_abs_2410_07073
institution arXiv
publishDate 2024
record_format arxiv
spellingShingle Pixtral 12B
Agrawal, Pravesh
Antoniak, Szymon
Hanna, Emma Bou
Bout, Baptiste
Chaplot, Devendra
Chudnovsky, Jessica
Costa, Diogo
De Monicault, Baudouin
Garg, Saurabh
Gervet, Theophile
Ghosh, Soham
Héliou, Amélie
Jacob, Paul
Jiang, Albert Q.
Khandelwal, Kartik
Lacroix, Timothée
Lample, Guillaume
Casas, Diego Las
Lavril, Thibaut
Scao, Teven Le
Lo, Andy
Marshall, William
Martin, Louis
Mensch, Arthur
Muddireddy, Pavankumar
Nemychnikova, Valera
Pellat, Marie
Von Platen, Patrick
Raghuraman, Nikhil
Rozière, Baptiste
Sablayrolles, Alexandre
Saulnier, Lucile
Sauvestre, Romain
Shang, Wendy
Soletskyi, Roman
Stewart, Lawrence
Stock, Pierre
Studnia, Joachim
Subramanian, Sandeep
Vaze, Sagar
Wang, Thomas
Yang, Sophia
Computer Vision and Pattern Recognition
Computation and Language
We introduce Pixtral-12B, a 12--billion-parameter multimodal language model. Pixtral-12B is trained to understand both natural images and documents, achieving leading performance on various multimodal benchmarks, surpassing a number of larger models. Unlike many open-source models, Pixtral is also a cutting-edge text model for its size, and does not compromise on natural language performance to excel in multimodal tasks. Pixtral uses a new vision encoder trained from scratch, which allows it to ingest images at their natural resolution and aspect ratio. This gives users flexibility on the number of tokens used to process an image. Pixtral is also able to process any number of images in its long context window of 128K tokens. Pixtral 12B substanially outperforms other open models of similar sizes (Llama-3.2 11B \& Qwen-2-VL 7B). It also outperforms much larger open models like Llama-3.2 90B while being 7x smaller. We further contribute an open-source benchmark, MM-MT-Bench, for evaluating vision-language models in practical scenarios, and provide detailed analysis and code for standardized evaluation protocols for multimodal LLMs. Pixtral-12B is released under Apache 2.0 license.
title Pixtral 12B
topic Computer Vision and Pattern Recognition
Computation and Language
url https://arxiv.org/abs/2410.07073