Molmo2

Training data disclosure: Partial list

Level determined for 18% of benchmark cells.

• TestTest data seen0
• SeenSeen, split unclear0
• TrainTraining data seen3
• Rel.Related data seen1
• NoNot seen0
• ?Unknown18

Built on

Everything these base models may have seen is inherited by Molmo2.

• Qwen3 · disclosure: Vague

  Molmo2-4B and Molmo2-8B (model card: Qwen/Qwen3-8B). Molmo2-O-7B uses OLMo (see training_data).

  Evidence

  “We release three versions of Molmo2: 4B and 8B models based on the Qwen3 LLMs [169], and a 7B model based on the OLMo LLM [112]”

• SigLIP 2 · disclosure: Partial list

  Model card text says SigLIP 2 but its link and base_model metadata point to google/siglip-so400m-patch14-384 (SigLIP 1); see open_questions.

  Evidence

  “All models use the SigLIP 2 So400m/14 384px ViT [139].”

Other declared training data

• pre-training data of base model OLMo 3 7B (Molmo2-O-7B only) (broad web data: may contain benchmark test sets) · split: unknown

  OLMo 3 is not in the allowed base-model id list for this pass, although olmo-3 exists in the registry.

  Evidence

  “Even with OLMo 3 as the LLM, our models still utilize a closed-data SigLIP 2 image”

• PixMo-Cap (captioning pre-training) (not yet matched to a dataset) · split: unknown
  Evidence

  “Our pre-training stage includes dense captioning with length conditioning and transcript prediction using PixMo-Cap”

• Tulu SFT data (English, no code) mixed into pre-training (not yet matched to a dataset) · split: unknown

  Tulu version not stated here.

  Evidence

  “We add NLP data using the supervised fine-tuning data from Tulu [71], filtered to remove non-English content and code”

• PixMo-Points, PixMo-Count, CoSyn-Point (pointing pre-training) (not yet matched to a dataset) · split: unknown
  Evidence

  “Additionally, we add pointing data from PixMo-Points, PixMo-Count, and CoSyn-Point [172].”

• SFT mixture: PixMo, nine new Molmo2 datasets, Tulu, open-source video and image datasets (not yet matched to a dataset) · split: unknown · fine-tuning
  Evidence

  “Our data mixture combines PixMo [29], the Molmo2 datasets, Tulu, and other open-source video and image datasets.”

• ChartQA · split: train of llm-multimodal · fine-tuning

  Reference [102] in the Image QA list is ChartQA. Split from the released SFT launch script (data loader split='train'); the paper itself only says 'following Molmo', and Molmo trained only on train sets.

  Evidence

  “open-source image datasets [48, 130, 101, 102, 103, 105, 65, 125, 94, 95, 14, 2, 61, 62, 107] following Molmo with CoSyn [172] instead of PixMo-Docs”

  “"chart_qa_weighted",”

  “split="train",”

• DocVQA · split: train of llm-multimodal · fine-tuning

  Reference [103] is DocVQA. Split from the released SFT launch script.

  Evidence

  “open-source image datasets [48, 130, 101, 102, 103, 105, 65, 125, 94, 95, 14, 2, 61, 62, 107] following Molmo with CoSyn [172] instead of PixMo-Docs”

  “"doc_qa",”

  “split="train",”

• AI2D (AI2 Diagrams) · split: train of llm-multimodal · fine-tuning

  Reference [65] is AI2D. Split from the released SFT launch script.

  Evidence

  “open-source image datasets [48, 130, 101, 102, 103, 105, 65, 125, 94, 95, 14, 2, 61, 62, 107] following Molmo with CoSyn [172] instead of PixMo-Docs”

  “"ai2_diagram_v2_mix_transparent",”

  “split="train",”

Benchmark cells

Classic LLM benchmarks

Multimodal LLM benchmarks

Recent LLM benchmarks

Notes

Paper notes Molmo2 lags on reasoning benchmarks (MathVista, MMMU), 'possibly due to the lack of multi-modal reasoning training data'. A decontamination step is reported only for MOSE videos against the MeViS validation set (video tracking), not for the multimodal benchmarks here. Accepts image input: “Molmo2 is a family of open vision-language models developed by the Allen Institute for AI (Ai2) that support image, video and multi-image understanding and grounding.”

Sources

• Molmo2: Open Weights and Data for Vision-Language Models with Video Understanding and Grounding (arXiv 2601.10611v4) · paper · retrieved 2026-10-15
• Molmo2-8B model card · model card · retrieved 2026-10-15
• allenai/molmo2 launch_scripts/sft.py · code · retrieved 2026-10-15
• Ai2 blog: Molmo 2 (2025-12-11; codebase released 3/3) · website · retrieved 2026-10-15