Llama 3.2-Vision (11B, 90B)

Training data disclosure: Vague

Level determined for 13% of benchmark cells.

• TestTest data seen0
• SeenSeen, split unclear0
• TrainTraining data seen0
• Rel.Related data seen0
• NoNot seen3
• ?Unknown19

Other declared training data

• 6B image and text pairs (not yet matched to a dataset) · split: unknown

  Source of the pairs not described.

  Evidence

  “Llama 3.2-Vision was pretrained on 6B image and text pairs.”

• Llama 3.1 text-only model (base language model) (not yet matched to a dataset) · split: unknown

  Inherits the Llama 3.1 language model and therefore its pre-training data; see llama-3.1.

  Evidence

  “Llama 3.2-Vision is built on top of the Llama 3.1 text-only model”

• publicly available vision instruction datasets (broad web data: may contain benchmark test sets) · split: unknown · fine-tuning
  Evidence

  “The instruction tuning data includes publicly available vision instruction datasets, as well as over 3M synthetically generated examples.”

• over 3M synthetically generated examples (not yet matched to a dataset) · split: unknown · fine-tuning
  Evidence

  “The instruction tuning data includes publicly available vision instruction datasets, as well as over 3M synthetically generated examples.”

• large-scale noisy (image, text) pair data (broad web data: may contain benchmark test sets) · split: unknown

  Same as the existing '6B image and text pairs' entry, described as noisy web-scale pairs.

  Evidence

  “First, we add image adapters and encoders, then pretrain on large-scale noisy (image, text) pair data.”

• medium-scale high quality in-domain and knowledge-enhanced (image, text) pair data (not yet matched to a dataset) · split: unknown

  Annealing stage (model card lists 'Stage 2 annealing').

  Evidence

  “Next, we train on medium-scale high quality in-domain and knowledge-enhanced (image, text) pair data.”

• synthetic QA generated with Llama 3.1 on in-domain images (not yet matched to a dataset) · split: unknown · fine-tuning
  Evidence

  “We leverage synthetic data generation by using the Llama 3.1 model to filter and augment question and answers on top of in-domain images”

• pre-trained image encoder (Llama 3 paper: ViT-H/14, MetaCLIP-style, 2.5B image-text pairs) (broad web data: may contain benchmark test sets) · split: unknown

  The Llama 3 paper describes the 'Llama 3-V' research models (8B/70B/405B); Meta does not explicitly state that the released 11B/90B Llama 3.2-Vision encoder is the same. Encoder is not in the allowed base-model id list.

  Evidence

  “we trained a set of adapter weights that integrate the pre-trained image encoder into the pre-trained language model.”

  “which has 630M parameters that were trained on 2.5B image-text pairs for five epochs.”

• academic datasets converted to QA pairs (image SFT data, Llama 3 paper) (not yet matched to a dataset) · split: unknown · fine-tuning

  Datasets not named; describes the Llama 3-V recipe in the Llama 3 paper, assumed (not stated) to match Llama 3.2-Vision. Overlaps the existing entry 'publicly available vision instruction datasets'.

  Evidence

  “We convert a highly filtered collection of existing academic datasets to questionanswer pairs using templates or via LLM rewriting.”

Benchmark cells

Classic LLM benchmarks

Multimodal LLM benchmarks

Recent LLM benchmarks

Notes

Modality set to text because the language backbone is evaluated on text benchmarks; the model also takes images. Built on the Llama 3.1 text model, so the llama-3.1 contamination findings may apply to its language backbone (not stated by Meta). Accepts image input: “Llama 3.2-Vision models come with multimodal (text and image) input capabilities enabling image reasoning applications.” Accepts image input: “The Llama 3.2-Vision instruction-tuned models are optimized for visual recognition, image reasoning, captioning, and answering general questions about an image.”

Sources

• Llama 3.2-Vision MODEL_CARD_VISION.md (meta-llama/llama-models) · model card · retrieved 2026-10-15
• meta-llama/llama-models README (launch dates) · readme · retrieved 2026-10-15
• Meta AI blog: Llama 3.2: Revolutionizing edge AI and vision with open, customizable models (2024-09-25) · website · retrieved 2026-10-15
• The Llama 3 Herd of Models (arXiv 2407.21783v3), Section 7 Vision Experiments · paper · retrieved 2026-10-15