Molmo
Allen AI's Open Vision-Language Model
Allen AI's open vision-language family — released September 2024 with sizes from 1B to 72B. Trained with a small (1M-image) high-quality dataset (PixMo), Molmo demonstrated that you don't need internet-scale image-text pretraining to get strong VLMs. Its signature trick: "pointing" — the model outputs (x,y) coordinates to point at things in images.
Cost
Free
Open weights — self-host
Context
4K
Up to 4,096 tokens
How are Intelligence, Speed & Cost bucketed?
Intelligence and Speed buckets are percentile ranks on
Artificial Analysis. Cost buckets are fixed dollar
thresholds keyed off output-token price ($/M out).
Intelligence
- Top 1%≤ 1%
- Top 5%≤ 5%
- Top 10%≤ 10%
- Good≤ 25%
- Medium≤ 50%
- Below avg> 50%
Speed
- Top 1%≥ 345 tok/s
- Top 5%≥ 237 tok/s
- Top 10%≥ 196 tok/s
- Good≥ 146 tok/s
- Medium≥ 90 tok/s
- Slow< 90 tok/s
Cost
- Freeopen weights · self-host
- Low< $1 / M out
- Moderate$1–5 / M out
- High≥ $5 / M out
Why it matters
Showed that high-quality small datasets + open weights beat the Llava-style "scrape and pretrain" approach for many VLM tasks. Molmo 72B held its own against Gemini 1.5 Pro and Claude 3 Opus at release with a fraction of the data.
Core Capabilities
Long Documents
Handles entire codebases, books, and multi-doc RAG.
Multimodal
Combines text, vision, and audio in one model.
Generative
Produces images, video, audio, or other media.
Vision
Understands images, scenes, and visual context.
Context Window
4k tokens
≈ short doc
4k This model 本模型
32k Long docs 长文档
128k Books 整本书
400k Multi-doc 多文档
1M Codebase 整个代码库
10M
Availability
API
Not available
Product / App
Not available
Open Source
Released
Enterprise
—
Pricing Model
Free / self-host
Open weights — pay only for compute.
Self-host Capability / Performance
Where this model sits relative to the middle 60% of models in the tree. All scores are 0–10 (higher is better).
Lower 20% Upper 80% This model
Context / memory
Context window size · log-scaled
6.0
9.0
1.0
Lower 20% 20th percentile — 20% of models score below this This model Where the current model lands Upper 80% 80th percentile — only 20% of models score above this
Percentile boundaries are computed across every model in the tree that reports the underlying benchmark for each capability.
What it feels like
- 72B variant outperforms GPT-4o (estimated 1T+ params) on image / chart / document understanding
- 7B variant comes close to GPT-4o-class performance — 'tiny model performs as well as powerful big ones' (MIT Tech Review)
- Trained on a curated 600K-image dataset (PixMo) — quality-over-quantity proof point
- All weights, training data, and inference code released — fully open, not just open-weights
- First open VLM that put Allen AI on equal footing with closed-source frontier
- Showed open-source vision is on par with proprietary in late 2024 — pivotal for the OSS narrative
Best use cases
- Self-hosted vision tasks at fraction of API cost
- Research and academic work needing full data + code access
- Distillation experiments and small-VLM training baselines
- Document QA, chart analysis, and image understanding without sending data out
Tools to try
Not ideal for
- Frontier visual reasoning by mid-2025 — Gemini 2.5 / 3 Pro and Claude 4 vision lead
- Edge / single-laptop deployments at the 72B scale
- Audio / video tasks (Molmo is image + text only at this generation)