Molmo 7B-D

Overview

Allen AI's Apache-licensed VLM built on Qwen2-7B and CLIP, scoring between GPT-4V and GPT-4o on benchmarks with unique pointing and grounding capabilities.

When to pick this model

• UI automation needing pixel-accurate pointing
• Visual grounding research with permissive licensing
• Image annotation pipelines requiring open data provenance
• Robotics and accessibility tools that need spatial references
• Replacing GPT-4V in workflows that demand on-prem deployment

VRAM requirements by quantization

Quantization	VRAM required
Q4_K_M (recommended)	5 GB
Q5_K_M	6 GB
Q8_0	9 GB
FP16 (no quantization)	16 GB

VRAM figures include model weights plus a typical 8k KV cache and ~600 MB runtime overhead (Ollama / llama.cpp baseline). Add headroom for higher context lengths.

Published benchmark scores

Benchmark	Score
MMMU	58.6

Scores published by the model author or aggregated from public leaderboards. Re-measured monthly by our editorial team.

Strengths

• Pointing capability is rare in open VLMs
• Apache 2.0 across weights and PixMo training data
• Performance lands between GPT-4V and GPT-4o on standard benchmarks
• Transparent human-annotated training set

Limitations

• 4096-token context cap limits multi-turn vision chats
• OCR quality trails Qwen2-VL 7B
• Smaller community ecosystem than mainstream VLMs

Architecture & training

Architecture: Dense · 7B vision · based on Qwen2 7B + OpenAI CLIP encoder

Training: AllenAI PixMo — original human pointing/annotation data, fully open.

Quick start

ollama run molmo

Or use the open-source MCP server to query this model from Claude Desktop, Cursor, or any MCP-compatible client.