Robot AI model
Magma
Magma is a multimodal agent model for visual understanding and goal-directed actions in digital interfaces and robot manipulation. It uses marked objects and motion traces during training; the paper reports 8.6B total parameters for Magma-8B and real-robot WidowX rollouts.
Records last verified Oct 13, 2026 · Send a correction
Also known as Magma-8B
- Organization
- Microsoft Research
- Type
- Foundation model
- Released / introduced
- Feb 18, 2025
- Public weights
- Public weights
- Licence / scope
- MIT
- Parameters
- 8.6B
- Base model
- LLaMA 3 8B with ConvNeXt visual encoder
- Humanoid association
- None reported
- Inputs
- vision, video, language
- Outputs
- text, visual action plans, robot actions, interface actions
Release dates can mark an announcement or paper introduction. Public checkpoints, code licences and weight terms are distinct; follow the publisher’s links for the stated scope.
Documented embodiments
Source-associated hardware; this list does not establish plug-and-play compatibility or commercial robot capability.
- WidowX
Real-robot manipulation rollouts on the project page.
Disclosed training data
- Open X-Embodiment with visual traces →
- Instructional video and interface-agent datasets
More robot AI models
Evidence
Sources
- [s1] Magma: A Foundation Model for Multimodal AI Agents ↗
Microsoft Research · paper · published Feb 18, 2025 · accessed Oct 13, 2026
- [s2] Magma-8B official model card ↗
Microsoft Research · official · accessed Oct 13, 2026
- [s3] Magma project: robot rollouts and training method ↗
Microsoft Research · official · accessed Oct 13, 2026
- [s4] Magma official repository and data releases ↗
Microsoft Research · official · accessed Oct 13, 2026
- [s5] Magma paper: model size and backbone ↗
Microsoft Research · paper · published Feb 18, 2025 · accessed Oct 13, 2026