Ai2Specialist modelOpen weights
Molmo2-4B-tb
Summary
Ai2 released Molmo2-4B-tb on November 6, 2026 as the high-level planner for ToyBlocks, a robotics environment for stacking and sorting blocks. The fine-tuned vision-language model reads a text goal and camera image, tracks progress, and decides what to do next.
It takes one 640 x 360 front-camera image as input and returns a short plan plus an action with a pointing target. A separate motion planner or learned action model such as ManiFlow-tb executes the movement. The card reports about 4.5 billion parameters and an operating rate of approximately 6 Hz.
The weights are released under Apache 2.0. Ai2 also notes that datasets used for the Molmo2-4B base model have terms for academic and non-commercial research. The card identifies research and education as the intended uses.
Similar models
More models of the same kind as Molmo2-4B-tb, chosen by model line, developer, size, price and release date.
- ManiFlow-tbAi2 · same developer
- Qwen-Drive-1.0-4BQwen · similar size
- Unlimited OCR 2Baidu · similar size
- FLUX 3 ActionBlack Forest Labs · similar size
- AstaBrief-8BAllen Institute for AI · similar size
- pplx-embed-v2-context-9bPerplexity · similar size
- pplx-embed-v2-late-9bPerplexity · similar size
- TeleOCRTeleAI · released around the same time
All specifications
Architecture
Watermarking and provenance
AI watermarking refers to invisible signals embedded in the output. Visible marks and provenance metadata are listed separately. Unknown means there is insufficient evidence for that particular field.
- AI watermarking
- Unknown
- no source
- Watermarking technique
- Not known
- no source
Licence and access
- License and use
- Apache 2.0. Ai2 notes that the base model was trained on datasets whose terms specify academic and non-commercial research use. The model card identifies research and education as the intended uses.
- Source: License and use
- Shutdown
- None announced
- no source
Further details
- Role
- High-level planner: reads a stacking goal and camera image, names the next step, and points to the relevant block or target. A separate low-level model controls movement.
- Source: Role
- Input and output
- One 640 x 360 front-camera image and a text goal as input. Output is a short plan with an action and Molmo2 target point.
- Source: Input and output