Molmo 72B

Positioning

Molmo 72B is the Allen Institute for AI's (AI2) 72B-parameter vision-language model with a unique capability: pointing. Unlike standard VL models that describe images in text, Molmo can identify specific regions within images — pointing to objects, people, text, or any visual element with pixel-level coordinates. For operators building interactive visual AI applications (visual Q&A with grounding, UI automation, accessibility tools), Molmo's pointing capability is a genuine architectural differentiator that no other open-weight VL model matches at this scale.

Strengths

• Unique pointing/grounding capability. Molmo can output pixel coordinates identifying specific regions in images — "point to the red car" returns coordinates. This is a first-class output modality, not a post-hoc add-on. No other 70B-class open-weight VL model offers this.
• AI2's research quality and transparency. As a non-profit research institute, AI2 releases detailed technical reports, training methodology documentation, and benchmark transparency. Operators who value understanding their model's internals will appreciate AI2's research-first approach.
• Competitive general visual understanding. Benchmarks on standard VL tasks (VQA, image description, document understanding) are competitive with InternVL 2.5 78B and Llama 3.2 90B Vision. The pointing capability is additive, not a tradeoff.
• Apache 2.0 license. The model card lists Apache 2.0 and says the model is intended for research and educational use. A genuine differentiator for operators who can't accept Llama or Mistral license restrictions.
• Strong performance on referring expression comprehension. Tasks like "find the object described by this text in the image" — where pointing is the natural output — are where Molmo's architectural advantage is most visible.

Limitations

• Pointing capability is powerful but narrow. For workloads that don't need visual grounding (standard image Q&A, description, classification), Molmo's unique advantage is unused and the model is a strong-but-not-class-leading VL model. The pointing capability is the reason to pick Molmo.
• Text-only performance trails dedicated LLMs. The 72B text backbone was optimized for VL alignment with pointing — text-only reasoning, coding, and general knowledge may trail similarly-sized text models like Qwen 2.5 72B Instruct.
• Ecosystem support for pointing output is nascent. Standard VL serving stacks (vLLM, Ollama) may not natively support coordinate output. Operators may need custom inference code to use the pointing capability.
• Base model license. Molmo is Apache 2.0, but the model card notes that the base model, Qwen2-72B, carries the Tongyi Qianwen license. Check it before commercial deployment.
• Published weights are large. The checkpoint is F32 with about 73B parameters, about 293 GB, and Ai2's card documents bfloat16 weights, about 147 GB. No GGUF of Molmo-72B is listed on Hugging Face and llama.cpp has no Molmo architecture, so Q4 and Q3 sizes do not apply.
• Pointing precision varies with image complexity. In cluttered images with many similar objects, pointing accuracy degrades. The model works best with clear, well-separated visual elements.

Real-world performance

• vs InternVL 2.5 78B: Molmo matches or approaches InternVL on general VL benchmarks while adding pointing capability that InternVL lacks. For pointing/grounding workloads, Molmo is the clear pick. For maximum general VL benchmark performance, InternVL may edge ahead.
• vs Llama 3.2 90B Vision Instruct: Molmo trades Llama's larger text backbone for pointing capability. For interactive visual AI applications (UI agents, accessibility tools, visual search), Molmo's pointing is the differentiator. For general VL tasks, Llama's text reasoning is stronger.
• vs Qwen 2.5-VL 72B: Both 72B VL models. Qwen 2.5-VL has stronger Chinese visual performance but uses the Qwen license. Molmo has Apache 2.0 and pointing capability. Pick by pointing need vs license/language preference.

Should you run this locally?

Yes if you're building interactive visual AI applications that require pointing/grounding (UI automation agents, visual Q&A with localization, accessibility tools for visually impaired users, visual search). Molmo's pointing capability is genuinely unique at this scale and enables applications that other VL models cannot handle.

No if your visual tasks don't need pointing (standard image Q&A or description) — pick InternVL 2.5 78B or Llama 3.2 90B Vision for stronger general VL benchmarks. Also no if the custom inference code requirement for pointing output is a blocker.

How it compares

• vs InternVL 2.5 78B: Molmo adds pointing; InternVL has stronger general VL benchmarks.
• vs Llama 3.2 90B Vision Instruct: Molmo adds pointing; Llama has larger text backbone and broader ecosystem.
• vs Qwen 2.5-VL 72B: Similar scale; Qwen has Chinese VL strength; Molmo has Apache 2.0 and pointing.
• vs Llama 3.3 70B: Molmo adds vision + pointing; Llama 3.3 is text-only. Different modality classes.

Run this yourself

• Datacenter: vLLM lists Molmo as a supported architecture. At bfloat16 the weights are about 147 GB, more than one A100 80GB holds.
• Vendor: huggingface.co/allenai/Molmo-72B-0924