What’s actually open: A field guide to China’s open-source world models

A reference inventory of the Chinese world models you can actually download, with licences read from the repositories rather than the press releases.

First compiled 23 July 2026. Updated 8 August 2026. This page is maintained — the field reorders itself monthly.

World models are the technology the robotics industry has decided will matter next, and China has spent the past eighteen months doing to them what it did to large language models: releasing the weights. That much has been widely observed. What has not been assembled anywhere is a straight answer to the obvious follow-up question — which ones, exactly, and on what terms.

This is that list. It covers models of Chinese origin that are open in the operative sense: a public repository, downloadable weights, and inference code you can run. It does not cover models that were announced, demonstrated, or published as papers. Every licence below was read from the repository’s licence file, not inferred from the word “open-source” in a company announcement — a distinction that turns out to matter more than expected, for reasons the third section explains.

Scope

Three tests for inclusion. The project must have a live official open entry point, normally a GitHub repository, with weights either in the repo or pointed to on Hugging Face or ModelScope. It must describe itself in official materials as a world model, world simulator, or world-action model, or publicly use one as a core engine. And it must originate in China or be led by a China-based team; joint international projects are marked as such.

Two exclusions worth naming. Closed models are out, including the two most famous: Google DeepMind’s Genie 3, shown only as a controlled preview, and World Labs’ Marble, sold as a hosted product. And NVIDIA’s Cosmos is out of the tables but belongs in the reader’s mind — it is open-weight under the Linux Foundation’s OpenMDW-1.1 licence, it is the only large-scale Western supplier of open world models, and any comparison that omits it is propaganda rather than inventory.

One structural warning before the tables. These projects are not commensurable. A model that generates a persistent 3D asset, a model that renders an interactive video stream, and a model that predicts a robot’s next action in latent space are three different products wearing one name. Ranking them against each other produces nonsense. They are grouped below by what they actually do.

Route 1 — Interactive worlds and 3D generation

Route 2 — Driving world models

Numerically the largest group — unsurprising, since autonomous driving structurally requires multi-step lookahead, counterfactual simulation and plan scoring. It is also, on the repository evidence, the least actively maintained. Six of the twelve entries below have had no repository activity for a year or more. Read this table as a record of what was opened, not of what is being worked on.

Route 3 — Embodied and world-action models

The direction of travel here is unification: not merely imagining future frames, but imagining, scoring, deciding and executing in one architecture.

The part the press releases don’t tell you: three licence regimes, not one

“Open-source” in this field describes at least three materially different offers, and the difference determines whether you can build a business on the model.

Permissive. Apache 2.0 or MIT, commercial use without asking. This covers most of the field: GigaAI, Daxiao, BAAI, Skywork, Motus, Vista, DriveDreamer, AgiBot (models only), Yume, Alibaba’s Wan. If you are choosing a foundation to build on, this is the tier that behaves the way “open-source” implies.

Encumbered. Open weights, non-commercial terms — usable for research, not for a product. Manycore’s SpatialLM carries CC-BY-NC-4.0 encoder weights. OpenDriveLab’s RISE splits the difference: Apache 2.0 code, CC BY-NC-SA 4.0 data and checkpoints.

Custom restrictive. Tencent’s world-model line — HunyuanWorld 1.0, HY-World 2.0, the HY-World 1.5 stack — ships under its own community licence, and the terms are specific: the agreement states on its first page that it does not apply in the European Union, the United Kingdom or South Korea, and any licensee whose products exceed one million monthly active users must return to Tencent for separate permission. The weights are genuinely downloadable. They are not unconditionally yours. Note that this is a per-line decision, not a company policy: Tencent’s embodied models take the opposite approach, with Hy-Embodied-RxBrain-1.0 released under Apache 2.0.

And one line moved. LingBot-World 1.0, released in January under Apache 2.0, is the most-starred Chinese world-model repository in existence. Its successor, LingBot-World 2.0 / Infinity, released on 9 July, ships under CC BY-NC-SA 4.0 — non-commercial, share-alike. Same company, same product line, six months apart, materially less open.

That single change is worth more attention than any benchmark result on this page. The prevailing story about world models is that China opens and the West closes. The repository record shows the picture is finer-grained than that: openness is a spectrum that individual companies move along, in both directions, and at least one of them has recently moved toward restriction on its flagship. Anyone treating “Chinese open-source” as a stable category should note that the category has an expiry date attached to specific model versions.

Two further cautions

Announcement dates are not release dates. Tencent announced HY-World 2.0 on 16 April; the release completed on 18 May. Daxiao’s Kairos 3.1 reached the repository on 2 July and was announced at WAIC on 19 July. GigaAI’s GigaWorld-1 topped WorldArena on 25 March; its weights repo was pushed on 12 July. Three companies, one pattern. Staged releases are normal engineering practice — but any inventory built from press coverage rather than repository logs will get its own timeline wrong.

Attention is concentrated in the corporate releases, and the gap is wide. Ranked by GitHub stars, the top of this field is almost entirely company work: LingBot-World (4,264), HunyuanWorld 1.0 (2,878), GigaBrain-0.1 (2,554), HY-World 2.0 (2,421), Matrix-Game (2,272), Kairos (2,220). The most-starred university or lab entry is Motus at 1,212, and most of the academic and driving projects sit in the low hundreds. This is not a judgement of research quality — DriveDreamer-Policy is the most complete model in its line and has 53 stars. It is a statement about where developer attention, and therefore ecosystem gravity, actually accumulates. The labs publish; the companies get built on.

There is no cross-domain ranking, and anyone offering one is selling something. The benchmarks in use here do not overlap: GameWorld Score for interactive game worlds, WorldNet for WorldGPT, FID/FVD and NAVSIM’s PDMS/EPDMS in driving, RoboTwin 2.0, LIBERO and WoWBench in robotics, WorldArena for embodied world models. Nearly every serious player leads on something. The question “which is China’s strongest open world model” has no answer until you specify the task domain, the output format and the benchmark — and even then the answer will have changed by the time you finish reading.

Method and corrections: every star count, licence and last-activity date above was checked against the repository on 23 July 2026 — via the GitHub API where it responded, otherwise from the repository page. “Last activity” is the most recent push to any branch, which is a rough maintenance proxy rather than a precise one: a dormant date is meaningful, a recent one may reflect a documentation fix rather than active development. Where a licence is given as “none declared”, the repository genuinely does not carry one; in at least one case (WorldGPT) a licence named in the README is not attached to the repo, which is a real ambiguity for anyone planning to build on it.