VLA GalleryAll modelsGallery

VLA Gallery / All models / Autoregressive & latent-action VLA

Magma

Microsoft Research·University of Maryland·UW–Madison·KAIST·University of Washington · 2025-02

VLAAutoregressive & latent-action VLA8.6Bdiscrete tokens

A multimodal agent foundation model that handles both UI navigation and robot manipulation.

View architecture diagram in the galleryPaper ↗

Paper

Magma: A Foundation Model for Multimodal AI AgentsarXiv 2502.13130

Overview

Category
Autoregressive & latent-action VLA
Type
VLA
Released
2025-02
Organization
Microsoft Research·University of Maryland·UW–Madison·KAIST·University of Washington
Open weights
Yes
Parameters
8.6B (paper)/9B (HF model card)

Architecture

Backbone LLM
LLaMA-3-8B
Vision encoder
ConvNeXt-XXlarge (CLIP)
Action head
256 discrete tokens (OpenVLA style) (discrete tokens)
Chunk
Not disclosed
Control rate
Not disclosed
Architecture (original summary)
LLaMA-3-8B + ConvNeXt-XXlarge (CLIP-ConvNeXt); robot actions as the last 256 discrete tokens of the vocabulary (OpenVLA style)

Training

Pretraining accelerators
NVIDIA H100·AMD MI300 (model card), count not disclosed
Pretraining duration
Not disclosed
Pretraining steps & batch
Up to 3 epochs, batch 1024, lr 1e-5, max sequence 4096
GPU/TPU-hours
Not disclosed
Fine-tuning
Not disclosed (training code released 2025-03-09)
Data type
Real-robot teleoperation + human video + web vision-language
Data scale
About 39M samples: robot OXE 970K trajectories (9.4M image-language-action), UI 2.7M, video 25M+ (about 4M clips), image-text 1.2M + Magma-SFT 820K
Main datasets
OXE, SeeClick, Vision2UI, Ego4D, Something-Something v2, Epic-Kitchens, ShareGPT4V, LLaVA-1.5
Tasks & evaluation
SimplerEnv 8 tasks (172 scenarios); real WidowX 4 tasks (about 50 trajectories per task); LIBERO (fine-tuned on 10 trajectories per task, 100 trials per suite); UI: Mind2Web·AITW
Robot
WidowX, SimplerEnv·LIBERO simulation, UI agent

Inference

Inference device & speed
Magma-8B bf16 1.1 s·17 GB per call, 4-bit 1.1 s·7 GB (README, GPU not stated)
Inference memory
bf16 17 GB · 4-bit 7 GB (README)
Weights bf16 (calc.)
17.2 GB
Where it runs
Not disclosed

Reinforcement learning & base

Reinforcement learning
Not used · Multimodal pretraining (supervised learning)
Base control
Fixed · No base control (arm only; UI agent is not a robot)

Notes

Key features
A multimodal agent foundation model that handles both UI navigation and robot manipulation. It grounds actions with Set-of-Mark and learns planning from video with Trace-of-Mark.
Notes
Training hardware appears only on the model card, not in the paper. Weights released 2025-02-25. Also covers UI agent tasks. Parameters (B) is the paper value 8.6B.
Verification
Confirmed

Other models in this category