Pixtral Large is a 124B open-weights multimodal model built on Mistral AI Large 2, with 69.4% on MathVista plus DocVQA and ChartQA results Mistral AI published at release, and a context window of 128K tokens that fits at least 30 high-resolution images.
import { streamText } from 'ai'
const result = streamText({ model: 'mistral/pixtral-large', prompt: 'Why is the sky blue?'})Released November 1, 2024, Pixtral Large is a 124B open-weights multimodal model built on Mistral AI Large 2. Pixtral Large's vision encoder carries one billion parameters, 2.5x larger than Pixtral 12B's encoder. The context window of 128K tokens accommodates at least 30 high-resolution images per request.
Pixtral Large scores 69.4% on MathVista. In Mistral AI's published evaluations at release, Pixtral Large's DocVQA and ChartQA scores were ahead of several proprietary multimodal models in the comparison set, including GPT-4o and Gemini-1.5 Pro. On the LMSys Vision Leaderboard, Pixtral Large led other open-weights models by approximately 50 ELO points. These results combine Mistral AI Large 2's text reasoning with the larger vision encoder's richer image representations.
Text-only performance stays comparable to Mistral AI Large 2, so Pixtral Large doesn't require a capability tradeoff when images are absent. Pixtral Large is available under the Mistral AI Research License for research and education, with a Mistral AI Commercial License for production use. Mistral AI has designated Pixtral Large as deprecated in favor of newer models.