Llama 3.2 90B Vision (multimodal)

À propos de Meta AI (FAIR)

Fondée 2013 · Menlo Park, California, USA

Meta AI is the research arm of Meta Platforms, established in 2013 as Facebook AI Research (FAIR) by Yann LeCun. FAIR has open-sourced many foundational models including PyTorch, RoBERTa, DETR, SAM and the LLaMA family. LLaMA 1 was released in February 2023, LLaMA 2 in July 2023, LLaMA 3 in April 2024 and LLaMA 3.1 (405B) in July 2024. Llama 3.2 launched in September 2024 at Meta Connect, introducing the first multimodal models in the LLaMA family (vision-enabled 11B and 90B) together with tiny on-device text-only siblings (1B, 3B). All Llama 3.2 vision weights are released under the Llama 3 Community Licence and are widely used by enterprise customers via Meta's partner ecosystem (Hugging Face, AWS Bedrock, Azure AI Studio, Google Vertex, Together AI, Groq, Fireworks).

Architecture

Decoder-only Transformer with cross-attended vision encoder

Llama 3.2 90B Vision combines the 70B-parameter Llama 3.1 text backbone (extended to 90B with vision components) and a Vision Transformer image encoder integrated via cross-attention adapter layers, similar in spirit to Flamingo but reusing the LLaMA architecture. The vision tower processes each image to a sequence of visual tokens which are injected into specific cross-attention layers of the LLM decoder while the original text-only weights remain frozen during the multimodal training stage, preserving text-only performance. Pretraining used 6B image-text pairs followed by multi-stage supervised fine-tuning and Direct Preference Optimisation (DPO) on a curated set of image instructions, math and chart data. The model supports a 128K context window and accepts up to 1120x1120 image inputs natively (with tiling for larger images). It does not support video or audio. Llama 3.2 90B Vision is released under the Llama 3 Community Licence (free for commercial use under 700M MAU).

Paramètres

90B

Contexte

128 000 tokens

Capacités

• Open-weights 90B vision-language model under Llama 3 Community Licence
• 128K token context window
• Image input up to 1120x1120 with tiling for larger images
• Chart, diagram, OCR and document understanding
• Strong on MMMU, MathVista, ChartQA and DocVQA among open-weights models
• Multilingual: English, German, French, Italian, Portuguese, Spanish, Hindi, Thai
• Tool use and JSON output via Llama 3.1 alignment recipe
• Best for: open-weights multimodal apps, on-premise document AI, indie research

Entraînement et licence

Pretrained on 6B image-text pairs from public web and licensed sources; supervised fine-tuning and DPO on curated multimodal instruction data. Text knowledge inherited from Llama 3.1 (15T tokens).

Licence: Llama 3 Community Licence: free for commercial use up to 700M MAU; redistribution must include the licence and acceptable use policy.

Tests de sécurité: Meta publishes a comprehensive model card with red-team findings on CBRN, child-safety and hate-speech vectors, plus Llama Guard 3 and Prompt Guard 2 companion models for production safety.