Mistral Pixtral Large (124B)

Acerca de Mistral AI

Fundado en 2023 · Paris, France

Mistral AI was founded in April 2023 in Paris by Arthur Mensch (CEO, ex-DeepMind), Guillaume Lample (Chief Scientist, ex-Meta FAIR LLaMA lead) and Timothee Lacroix (CTO, ex-Meta FAIR). The company is the European poster child for sovereign AI and has raised a record-setting $113M seed (June 2023) followed by $415M Series A (December 2023), $645M Series B (June 2024) and a reported €1B+ funding round in 2025 at a valuation above $6B, with backers including Andreessen Horowitz, Lightspeed, General Catalyst, BPI France and Cisco. Mistral's model line includes Mistral 7B, Mixtral 8x7B and 8x22B (MoE), Mistral Large 1/2, Codestral, and the multimodal Pixtral family. Pixtral 12B launched in September 2024 and Pixtral Large 124B followed in November 2024 as the flagship multimodal model, released under the Mistral Research Licence and offered commercially via la Plateforme.

Arquitectura

Decoder-only Transformer with Pixtral-ViT vision encoder

Pixtral Large is a 124B-parameter multimodal model built on top of Mistral Large 2 (123B decoder-only Transformer) and Mistral's custom Pixtral-ViT vision encoder (1B parameters, 400M visible). The vision encoder processes each image at its native aspect ratio in patches of 16x16 pixels and projects the resulting visual tokens directly into the LLM token stream (without cross-attention or fixed-size resampler), which lets the model handle high-resolution images and many images per prompt in the same 131K-token context window. The vision encoder uses 2D rotary position embeddings and an attention mask that distinguishes patches from different images. Training used a multi-stage curriculum: vision-encoder pretraining on image-text pairs, joint vision-language pretraining and multimodal supervised fine-tuning with chain-of-thought style instruction data. Pixtral Large posts top scores on MathVista, DocVQA and VQAv2, often matching GPT-4o and Claude 3.5 Sonnet at lower cost. The model is released under the Mistral Research Licence with a separate commercial licence available.

Parámetros

124B (123B LLM + 1B vision encoder)

Contexto

131.072 tokens

Capacidades

• Native variable-resolution image input (no fixed grid)
• Up to 30 images per request, 131K-token context
• Top-tier MathVista, DocVQA, VQAv2 and ChartQA scores
• 131K-token context window
• Multilingual: English, German, French, Spanish, Italian, Portuguese, Dutch, Russian, Arabic, Hindi, Japanese, Chinese, Korean
• JSON output and tool use inherited from Mistral Large 2
• Open weights under Mistral Research Licence (commercial via separate licence)
• Best for: high-end open-weights multimodal apps, document AI, math-with-image, European data sovereignty

Entrenamiento y licencia

Mistral has not disclosed exact dataset sizes. The technical report describes a 'multimodal corpus of public web data, licensed image-text pairs and curated instruction data' with a knowledge cutoff in 2024.

Licencia: Weights under Mistral Research Licence (non-commercial). Commercial use via Mistral la Plateforme API or a paid Mistral Commercial Licence.

Pruebas de seguridad: Mistral publishes the Pixtral technical report with red-team findings on hallucination and bias; the model is integrated with Mistral's safety filters on la Plateforme.