Allen AI's Molmo2-ER Beats GPT-5 and Gemini on Robot Vision at 4B Parameters

Allen AI has quietly shipped a small vision language model that punches well above its weight for robotics work. Molmo2-ER is a 4B parameter model tuned for the perception skills robot policies actually need: pointing at exact pixels, matching objects across camera views, and reasoning about geometry over video. It is Apache 2.0 licensed, available on Hugging Face, and works out of the box with Transformers, vLLM, and SGLang.

A 4B model that outscores the giants

Molmo2-ER outperforms every open-weight baseline along with the strongest closed-source models, including Gemini Robotics-ER 1.5 Thinking and GPT-5, on 9 of 13 established embodied reasoning benchmarks. It posts an overall average of 63.8%, a 17-point jump over the Molmo2 starting point.

On the same 13-benchmark suite (Point-Bench, RefSpatial, BLINK, CV-Bench, ERQA, EmbSpatial, MindCube, SAT, VSI-Bench, and others), the comparison looks like this:

• Molmo2-ER: 63.8
• Gemini Robotics-ER 1.5 Thinking: 61.3
• Qwen3-VL-8B: 61.0
• GPT-5: 57.9
• Gemini 2.5 Pro: 57.1

Why a specialized backbone was needed

Existing VLM backbones are optimized for semantic image understanding rather than the metric, geometric, and temporally grounded reasoning required for robot control. A general VLM can tell you there is a mug on a table, but it struggles to say exactly which pixel to grasp, how far away the object sits, or which item in camera B corresponds to the one in camera A.