VLA Gallery / All models / New in 2026

Being-H0.7

BeingBeyond · 2026-04

기타New in 20263BNot disclosed

A 3B model introduced by BeingBeyond as its 'flagship WAM', reporting a LIBERO average of 99.2%.

Paper

Being-H0.7: A Latent World-Action Model from Egocentric VideosarXiv 2605.00078

Overview

Category

New in 2026

Type

기타

Released

2026-04

Organization

BeingBeyond

Open weights

Unverified

Parameters

3B

Architecture

Backbone LLM

InternVL3.5 (understanding) + Qwen3 (action expert)

Vision encoder

—

Action head

Unverified (Not disclosed)

Chunk

Not disclosed

Control rate

Not disclosed

Architecture (original summary)

InternVL3.5 understanding expert + Qwen3 action expert; action representation·chunk unverified

Training

Pretraining accelerators

Unverified

Pretraining duration

Unverified

Pretraining steps & batch

Unverified

GPU/TPU-hours

Not disclosed

Fine-tuning

Unverified

Data type

Unverified

Data scale

Unverified

Main datasets

Unverified

Tasks & evaluation

LIBERO average 99.2%; RoboCasa-50 62.1%; CALVIN 4.67·4.48

Robot

Unverified

Inference

Inference device & speed

Unverified

Inference memory

Not disclosed

Weights bf16 (calc.)

6 GB

Where it runs

Not disclosed

Reinforcement learning & base

Reinforcement learning

Not disclosed · Training method unverified

Base control

Fixed · Outputs upper-body actions only (G1 26 DoF upper-body interface); lower body·waist generated by a separate balance WBC module, no mention of locomotion

Notes

Key features

A 3B model introduced by BeingBeyond as its 'flagship WAM', reporting a LIBERO average of 99.2%. Accepted to NeurIPS 2026.

Notes

New facts from verification notes (NEW). README announcement 2026-04-14 ('flagship WAM'), arXiv 2605.00078 (abs date not read), accepted to NeurIPS 2026 (announced 2026-09-24). Paper figures are the paper summary values in the verification notes. Type follows the company's own label (WAM).

Verification

Confirmed