VLA GalleryAll modelsGallery

VLA Gallery / All models / Industry models

GEN-0

Generalist AI · 2025-11

기타Industry models1B · 6B · 7B comparisonNot disclosed

An embodied foundation model pretrained on 270,000 h of real-world manipulation data.

View architecture diagram in the galleryOfficial page ↗

Paper

No paper

Overview

Category
Industry models
Type
기타
Released
2025-11
Organization
Generalist AI
Open weights
No
Parameters
1B · 6B · 7B comparison; scaled up to 10B+

Architecture

Backbone LLM
Not disclosed
Vision encoder
—
Action head
Harmonic Reasoning (asynchronous continuous-time token stream) (Not disclosed)
Chunk
Not disclosed
Control rate
Not disclosed
Architecture (original summary)
Not disclosed; 'Harmonic Reasoning' handling asynchronous continuous-time streams of perception·action tokens (thinking and acting simultaneously)

Training

Pretraining accelerators
Not disclosed (only data processing O(10K) cores disclosed)
Pretraining duration
Not disclosed
Pretraining steps & batch
Not disclosed
GPU/TPU-hours
Not disclosed
Fine-tuning
Not disclosed (SFT on 16 task sets simultaneously; post-training compared with task data from 5.6 h (1%) to 550+ h (full))
Data type
Real-world manipulation data (robots·collection devices; equipment details not disclosed)
Data scale
270,000 h, over 10,000 h added per week; thousands of homes, warehouses and workplaces, thousands of collection devices and robots; absorbs 6.85 years' worth per day of training
Main datasets
In-house collected data (private)
Tasks & evaluation
Clothes handling, LEGO assembly, camera·phone·box packing etc.; up to 99% on some tasks
Robot
6DoF, 7DoF, 16+DoF semi-humanoid robots

Inference

Inference device & speed
Not disclosed
Inference memory
Not disclosed
Weights bf16 (calc.)
20 GB
Where it runs
Not disclosed

Reinforcement learning & base

Reinforcement learning
Not disclosed · Training method not disclosed
Base control
Fixed · No mention of base·locomotion control (6·7 DoF arms, 16+DoF semi-humanoid)

Notes

Key features
An embodied foundation model pretrained on 270,000 h of real-world manipulation data. Reports a phase transition at 7B, with smaller models showing 'ossification', and a power law between pretraining data scale and downstream performance.
Notes
Parameters (B) is 10, the lower bound of '10B+'. Uses language-conditioned data in post-training, but the company does not call it a VLA, so type is Other. Data scale is a claim made only on the company blog (no external verification). Announced 2025-11-04.
Verification
Confirmed

Other models in this category