Being-H: a family of human-centric embodied models split across WAM and VLA lines

Two model lines, not one, and the repository mirrors that

Being-H is a family rather than a single model, and the split is by architecture. Inside the repository, Being-H0.7 is the flagship WAM model and Being-H0.5 is the flagship VLA model. The top level of the tree reflects that split directly, with a `Being-H05/` directory, a `Being-H07/` directory, a `tutorials/` directory, the `LICENSE` file, and the `README.md`.

The model family table positions three entries. Being-H0.7 is the flagship WAM, described as a latent world-action model from egocentric videos with future-aware latent reasoning. Being-H0.5 is the flagship VLA, a human-centric VLA model for cross-embodiment generalization with a unified action space. Being-H0 sits on a separate `being-h0` branch and is the previous VLA, the first Being-H release for human-video VLA pretraining.

Being-H0.5 and Being-H0 both link to model collections on Hugging Face, while Being-H0.7 links only to a blog and a paper. That difference is the single most important practical fact in the table, and the news section makes it explicit. On 2026-04-14 the team published Being-H0.7 and wrote that code and checkpoints are coming soon. The 2026-09-24 entry says Being-H0.7 is accepted by NeurIPS 2026, but acceptance is not a release of weights.

So the practical reading of the repository is: two VLA lines with published artefacts, one WAM line with a paper and a promise, and a tutorial workspace that runs against the older flagship.

The release timeline tells you what is actually usable

The dated news list is specific, which makes it a better source of truth than the marketing framing. The earliest entry is 2025-07-21, when the team published Being-H0, described as the first human-video VLA release, with a paper link. On 2025-08-02 the Being-H0 codebase and pretrained models were released through the BeingBeyond Hugging Face collections.

Being-H0.5 followed on 2026-01-20 as the flagship VLA model for cross-embodiment generalization, and on 2026-01-24 the team updated the H0.5 training, inference, and data preparation docs and open-sourced post-training data for PND Adam-U through a Hugging Face dataset collection. That January documentation update is the moment the model became something a newcomer could pick up, because a model without a data preparation guide is a paper with a download button.

The data side opened up in stages. On 2026-03-20 the UniHand_Preview dataset was released as a subset of the Being-H0.5 pre-training mixture, and it is linked on Hugging Face as a dataset rather than a model collection.

Acceptance news follows the releases rather than leading them. Being-H0 was accepted by ICML 2026 on 2026-05-01, Being-H0.5 by CoRL 2026 on 2026-09-05, and Being-H0.7 by NeurIPS 2026 on 2026-09-24. On 2026-06-09 the team added Being-H-EDU, an educational tutorial workspace for post-training and deployment examples.

There are no GitHub releases attached to the project, so the tags and changelog are not there to consult. The dated news list in the project documentation is the version history.

The tutorial workspace is the intended entry point

The `tutorials/` directory collects practical examples for adapting, post-training, evaluating, and deploying Being-H models on educational robots and community benchmarks. Inside it, `tutorials/Being-H-EDU/` is an educational workspace for data processing, post-training, and local robot deployment with Being-H0.5.

The scope is stated with a useful limit: the current public example supports SO101. That single sentence tells you what kind of project this is aimed at. SO101 is the robot named as the current public example, and the tutorial path is data processing through post-training through local deployment on one specific machine. This is a research and teaching path, not a production robotics stack.

There is no install command in the repository, no environment specification file at the root, and no wheel or container published under the Being-H name. The route to running the code is therefore: pick a model line, fetch the checkpoints and data from the linked Hugging Face collections, and follow the training, inference, and data preparation docs that the January 2026 update refreshed. The research site at research.beingbeyond.com hosts the per-release blog and paper pages, including a PDF for Being-H0.7, which is where the method details live.

That is an unusual shape for a repository. Instead of a self-contained code release, you get a hub of subprojects with per-model documentation and external artefacts, which suits a research lab publishing several related lines over time and suits a newcomer less.

A family with downstream projects already on top of it

The project maintains a list of work built on the Being-H family, which is the best available evidence of whether the models are usable by outside teams. Seven entries are listed, each with a paper, a website, and for most of them a GitHub repository.

Two of them attack the same problem from opposite sides. TTP is human-centric transferable tactile pre-training for dexterous robotic manipulation, and PTR is conservative offline robot policy learning via posterior-transition reweighting. A third, JALA, is joint-aligned latent action, aimed at scalable VLA pretraining in the wild, and VIPA-VLA is spatial-aware VLA pretraining through visual-physical alignment from human videos.

Then there are the critical ones. BeTTER is titled Unmasking the Illusion of Embodied Reasoning in Vision-Language-Action Models, which is a challenge to the field rather than an extension of it, and Rethink_VLA is titled Rethinking Visual-Language-Action Model Scaling: Alignment, Mixture, and Regularization. Both live under the BeingBeyond organisation and both interrogate assumptions in the area the family targets.

One entry, DexHiL, a human-in-the-loop framework for vision-language-action model post-training in dexterous manipulation, linked to its own project page. The list ends with an invitation to open a pull request to share work built on Being-H.

Read as a set, these projects cover pretraining data, tactile sensing, offline policy learning, post-training with a human in the loop, and an evaluation that questions whether the reasoning is real. That is a reasonable spread for a model family that has been public since mid 2025, and it tells you the intended use is as a starting point for research rather than a finished product.

Licence terms are simple, but the model terms live elsewhere

The repository itself is released under Apache-2.0, with the licence in the `LICENSE` file at the root. That covers the code in the tree, which is the `Being-H05/`, `Being-H07/`, and `tutorials/` directories.

What the repository licence does not settle is the terms attached to the checkpoints and datasets, because those live on Hugging Face rather than here. The Being-H0 and Being-H0.5 model collections, the PND Adam-U post-training data collection, and the UniHand_Preview dataset are separate artefacts with their own terms and their own model cards, and anyone shipping a fine-tuned checkpoint needs to read the card for the specific base model rather than assume the Apache-2.0 grant in the repository extends to it.

There is also the question of what a downstream product inherits from the papers. Being-H0.5 and Being-H0.7 are arXiv preprints, and Being-H0 is listed as an ICML 2026 proceedings paper. The citations given in the project are ordinary academic citations with author lists and venues, and they are the mechanism by which prior work in this area is attributed.

A last practical point on upkeep. The last push to the repository was on 2026-09-27, so the tree is being touched, but with no GitHub releases and no changelog, an upgrade is whatever the current state of `main` happens to be. For a model family this is normal, and for anything you intend to reproduce it means you should pin a commit and record it, rather than tracking a branch.

Where Being-H is the wrong tool

The limits here are structural rather than bugs.

First, this is a research family built around human video pretraining and cross-embodiment generalisation. If your problem is a single robot with a single task and a well defined action space, the unified action space and the world-action modelling are overhead. The claims in the model table are about generalisation across embodiments, and nothing in the repository suggests a task-specific quick start.

Second, the flagship model is the one you cannot download. Being-H0.7 is described in the model family table as the flagship WAM, and the release note says code and checkpoints are coming soon. As of the last news entry about it, on 2026-09-24, it has a paper and an acceptance and no published artefacts. If your plan depends on the WAM line, the plan depends on a future release.

Third, the deployment story stops at an educational arm. The tutorial workspace supports SO101, and the words used are educational robots and community benchmarks. There is no simulation-only quick path documented, and no indication that the models run on a laptop without a GPU, so anyone without robotics hardware is looking at a much longer path than the tutorial suggests.

Fourth, this is a research artefact with a moving target. The project publishes checkpoints, datasets, and docs for two different model lines and a third in waiting, and the tutorial examples are pinned to a specific model. Following a blog post from 2026 with a repository as of late 2026 is a real coordination cost, and that cost is the main practical argument for choosing a more settled base model.

Editorial conclusion

Being-H suits a research team that wants to pretrain or post-train vision-language-action policies on human video and then port them to a small educational robot, and it does not suit a team looking for a packaged, pip-installable library with a stable API. Before building on it, check three things: which of the three releases you actually need, since Being-H0.7 was announced with code and checkpoints still to come and Being-H0.5 and Being-H0 are the ones published through Hugging Face, whether your hardware is one of the community benchmarks the tutorials target, with SO101 named as the current public example, and whether the Apache-2.0 licence and the per-model paper terms fit your intended use.

Frequently asked questions

Official sources

1. BeingBeyond/Being-H on GitHub
2. Issues
3. License: Apache-2.0
4. Project website
5. README