Ranked by source signals (HN, Lobsters, authority domains, official posts) and recency. Summaries are generated from a stored copy of the article text; every sentence cites a span of that text, and nothing is shown that failed the citation check.
Official3
molmo.allenai.org2024-09-25officialauthorityabout the Molmo familyHN 55 pts · 10 comments2,409 words
Ai2 released Molmo, a family of open multimodal AI models that outperform proprietary systems like GPT-4o on benchmarks. The models utilize a new dataset called PixMo, which relies on high-quality human annotations rather than synthetic data from closed models.written, reviewed, cited
- Molmo trains on less than 1M image text pairs, representing 3 orders of magnitude less data than many competitive approaches.
- Annotators describe images in speech for 60 to 90 seconds rather than writing descriptions to avoid copying proprietary VLM outputs.
- The PixMo-Points dataset contains 2.3 million question-point pairs from 428,000 images to enable models to point to visual elements.
molmo.allenai.org2024-10-23officialauthorityabout the Molmo familyHN 4 pts · 1 comments317 words
Ai2 Playground is a free scientific and educational tool allowing interaction with the Olmo 3.1 32B Instruct model, which primarily handles English queries. Users can enable tool calling and view OlmoTrace results showing training documents with exact text matches. Ai2 may use uploaded content to train, evaluate, and improve AI models, with optional consent to publish de-identified conversations in open datasets.written, reviewed, cited
allenai.org2025-12-16officialauthorityabout the Molmo familyHN 1 pts · 1 comments2,137 words
AllenAI released Molmo 2, an open multimodal model family for video understanding that trains on 9.19M videos versus 72.5M used by Meta's PerceptionLM, representing less than one-eighth of the data. The models support inference with sequence lengths up to 16,384 tokens and sample up to 128 frames at ≤2 fps per clip during training. Three variants are available, including a Qwen 3-based 8B model and a fully open Olmo-backed 7B option.written, reviewed, cited
- Molmo 2 answers 'how many?' questions by returning concrete visual evidence through spatial and temporal grounding, not just a number.
- The model maintains a stable ID for a person across a sequence to avoid double-counting when analyzing entire clips for specific events.
- Training uses a mix of 60% captioning, 30% pointing, and 10% natural language data to preserve strong language capabilities during pretraining.
Analysis4
github.com2026-03-04authorityabout the Molmo familyHN 2 pts · 0 comments2,802 words
Molmo2 is an open-source vision-language model from Ai2 that excels at point-driven grounding across single images, multi-images, and video. The repository requires python >= 3.11 and supports training with sequences up to 36k+ tokens and 384 frames. Models utilize backbones including qwen3_4b_instruct, qwen3_8b, and olmo3_7b_instruct.written, reviewed, cited
- Some datasets like MLVU and LongVideoBench require accepting HuggingFace user agreements before download scripts function.
- The Long-Context SFT checkpoint format differs slightly from Hugging Face repo checkpoints despite being backwards-compatible.
- Context Parallelism shards long sequences across multiple GPUs to enable training on videos that would not fit in single GPU memory.
interconnects.ai2024-09-28authorityabout the Molmo familyHN 2 pts · 0 comments2,874 words
Meta and Ai2 released late-fusion multimodal models: Llama 3.2 Vision in 1B, 3B, 11B-Vision, and 90B-Vision sizes under the Llama 3.2 community license agreement, while Molmo variants use the Apache 2.0 license. Molmo was trained on millions of primarily human datapoints using a CLIP encoder, whereas Llama 3.2 V utilized approximately 6B (image, text) pairs, mostly synthetic data. Molmo distinguishes itself with the ability to point at pixels in referenced images, a capability absent in its peers.written, reviewed, cited
- Molmo models are fine-tuned from a non-open-source model and updating parameters prevents them from being classified as open source.
- Molmo training involved annotators responding to images in audio rather than text, enabling far more creative and descriptive annotations.
molmo.org2024-09-26about the Molmo familyHN 4 pts · 0 comments
wired.com2024-09-25about the Molmo familyHN 1 pts · 0 comments697 words
