China's open-source AI push — and what it means for humanoids

DeepSeek's V3 (December 2024) and R1 (January 2025) started the wave. V3 is a Mixture-of-Experts (MoE) model with 671B total parameters, of which 37B are activated for each token. DeepSeek says pre-training took 2.664M H800 GPU hours. R1 ships under the MIT License, and its card says it allows 'any modifications and derivative works, including … distillation for training other LLMs.'

By late 2026 the large Chinese labs all publish frontier-scale weights on Hugging Face. They are DeepSeek (V4-Pro, 1.6T backbone parameters with 49B activated; V4.1-Flash, 552B backbone), Alibaba Qwen (Qwen3.8-2.4T-A95B), Moonshot (Kimi K3, 2.8T total with 104B activated), Zhipu/Z.ai (GLM-5.3), MiniMax (M3, about 428B), StepFun (Step 5 Preview, 600B total with 27B active) and Baidu (ERNIE 5.5 Flash Base, 280B total with 14B activated). Tencent publishes Hunyuan models, but at smaller sizes such as Hunyuan-A13B (80B total, 13B active).

Hugging Face's Summer 2026 report says that 'in almost every month of 2026, the largest and most performant open model from a Chinese lab was larger than any model an American lab released.' It puts China's monthly ceiling between 754B and 2.78 trillion parameters.

Hugging Face counted 178 Chinese releases above 20B parameters in 2026. Of those, '59% carry Apache 2.0 and 22% carry MIT, and almost none carry non-commercial restrictions.' DeepSeek and Z.ai ship models between 700B and 1.65T parameters under plain MIT. DeepSeek V4-Pro-0813 and V4.1-Flash are MIT, and Qwen3.8-27B is Apache-2.0.

The largest releases now carry custom licences. The Kimi K3 License requires a separate agreement with Moonshot when a Model-as-a-Service operator has more than US$20 million in revenue over 12 months. It also requires 'Kimi K3' to be shown in products with more than 100 million monthly active users (MAU) or more than US$20 million in monthly revenue. The Qwen3.8-Max License (on the 2.4T model) requires a separate licence for Model-as-a-Service or 'AI Work Assistant' businesses above US$50 million in 12-month revenue. The Qwen Community License 1.0 (on Qwen3.8-Flash-Next) requires a licence for those businesses at any revenue level.

The MiniMax Community License on M3 grants rights 'for non-commercial purposes'. Commercial users must display 'Built with MiniMax M3', and companies with more than US$20 million in yearly revenue need prior written authorisation. The licence also bars military use. The GLM-5.3 License requires a Z.ai security review only for Model-as-a-Service operators with more than US$10 billion in 12-month revenue. Tencent's Hunyuan licence 'does not apply in the European Union, United Kingdom and South Korea.'

Hugging Face's Spring 2026 report found that Chinese models accounted for 41% of downloads on the Hub over the past year. It also found that Alibaba alone had more derivative models than Google and Meta combined, with more than 113,000 Qwen-family derivatives.

By the Summer 2026 report, 'Qwen-based models now account for 151,448 derivatives on the Hub, 2.6× Meta's total footprint and 4.7× the Llama repositories specifically.' New Qwen derivatives appeared at about 180–210 a day through the first seven months of 2026.

At the March 2026 National People's Congress, Minister of Industry and Information Technology Li Lecheng said Chinese firms' open-source large models 'rank first globally in downloads'. Xinhua cited OpenRouter data showing that calls to Chinese models began surpassing U.S. levels in February 2026.

The State Council's 'AI+' Opinions (August 2025, in CSET's translation) contain a section titled 'Promote a thriving open-source ecosystem.' It calls for supporting open-source AI communities, for 'the gathering together and openness of models, tools, and datasets', for counting open-source contributions toward university credit, and for 'internationally influential open-source projects.' The same document sets adoption targets for smart terminals and AI agents: above 70% by 2027 and above 90% by 2030.

In January 2026, eight departments led by MIIT issued an AI-plus-manufacturing plan that aims to 'build a world-leading open-source ecosystem' by 2027. The 2026 Government Work Report 'emphasizes strengthening open-source AI ecosystems.' According to Xinhua, the 15th Five-Year Plan (2026–2030) outline aims to 'accelerate the building of an open-source technology ecosystem and community with global reach' and names embodied AI as a field to advance.

MIIT's humanoid-robot Guiding Opinions (dated October 2023) call for a large-model-based robot 'brain'. By 2025 they target breakthroughs in the 'brain, cerebellum and limbs', with mass production. By 2027 they target a secure, reliable supply chain at a world-advanced level. In July 2026 the NDRC and other departments released an AI cooperation action plan calling for 'broader sharing of open-source AI ecosystems.'

U.S. export controls are part of the background. On 13 January 2026, BIS moved license applications for the Nvidia H200, AMD MI325X and similar chips bound for China to case-by-case review. Applicants must meet conditions, including independent third-party testing in the United States. Hugging Face's Spring 2026 report notes that Chinese open models 'are being released with explicit support for domestically developed chips'.

Prices are low. DeepSeek's API lists deepseek-flash (V4.1-Flash-1017) at US$0.30 per 1M input tokens (cache miss) and US$1.20 per 1M output tokens at peak, and half that off-peak. deepseek-v4-pro is US$1.32 input and US$3.96 output at peak. Xinhua said Chinese models held three of OpenRouter's top five usage spots, 'mainly driven by their high cost-performance.'

For humanoid makers, permissive weights mean a robot's language and vision 'brain' can be fine-tuned and run on-device without per-call fees, subject to each licence's revenue thresholds. Qwen's model cards include embodied-reasoning tests. Qwen3.8-Flash-Next reports an 'Embodied intelligence (ERQA)' score. Qwen-Drive-1.0-4B is an Apache-2.0 model built on Qwen3.5-4B that combines 3D perception and motion planning for driving. Hugging Face reports that repositories using lerobot (its robot-learning library) grew 194% in 2026.

Since late 2024, Chinese humanoid companies and state-backed labs have released four kinds of robot software publicly: large robot-demonstration datasets, vision-language-action (VLA) models that turn camera images and instructions into motor actions, video 'world models' that predict what happens next and can act as simulators, and coordination software described as a robot operating system. Code, and usually weights, are posted on GitHub and Hugging Face, often with a ModelScope mirror.

The best-known releases come from AgiBot (AgiBot World dataset, GO-1, Genie Envisioner), Unitree (UnifoLM-WMA-0, UnifoLM-VLA-0), the Beijing Academy of Artificial Intelligence (RoboBrain 2.0, RoboOS), the Beijing Humanoid Robot Innovation Center / X-Humanoid (RoboMIND, Pelican-VL, XR-1), Alibaba's DAMO Academy (RynnVLA-001, RynnBrain), Galbot with Peking University (GraspVLA, SynGrasp-1B) and Fourier (N1 open hardware).

AgiBot released AgiBot World Alpha on 2024-12-30 and AgiBot World Beta on 2025-03-01. The repository describes Beta as the complete dataset of 1,003,672 trajectories (~43.8T) and Alpha as a curated subset of 92,214 trajectories (~8.5T). The technical report (arXiv 2503.06669) says the data came from 100 real robots across 217 tasks in five deployment scenarios: domestic, retail, industrial, restaurant and office.

GO-1 ('Genie Operator-1'), a generalist robot policy trained on that dataset, was open-sourced on 2025-09-19. Its Hugging Face checkpoint holds about 2.82 billion BF16 parameters, and the paper says it uses InternVL2.5-2B as its vision-language backbone. The repo points to Genie Studio, which ships the AgiBot G1 with GO-1 out of the box.

AgiBot's Genie Envisioner world-model platform released its code on 2025-08-13 and GE-Base weights on 2025-08-14. GE-Sim weights based on Cosmos2 followed on 2025-10-22, Sim v2.0 on 2026-05-28 and Act v2.0 on 2026-09-10.

AgiBot's releases are non-commercial. The GO-1 repo states that all data and code are under CC BY-NC-SA 4.0, and Genie Envisioner uses the same licence except for code adapted from Apache-2.0 projects.

The Beijing Academy of Artificial Intelligence (BAAI) released RoboBrain 2.0, an 'embodied brain' model for perception, reasoning and planning. The 7B checkpoint came out on 2025-06-06, the 32B on 2025-07-03 and the 3B on 2025-07-23, all under Apache-2.0. Downloading requires accepting the terms on Hugging Face. BAAI's model card claims it outperforms Gemini 2.5 Pro, o4-mini and Claude Sonnet 4 on several spatial and temporal benchmarks. That is the developer's own claim.

BAAI's RoboOS (Apache-2.0, technical report arXiv 2505.03673) describes itself as 'the first open-source embodied operating system'. It has a 'brain–cerebellum' design: a cloud multimodal model handles planning, a plug-in skill library executes actions, and shared memory coordinates several robots.

The Beijing Humanoid Robot Innovation Center (X-Humanoid, founded November 2023) publishes under the X-Humanoid organisation on Hugging Face. Its releases include the RoboMIND dataset (arXiv 2412.13877): 107k real-world trajectories over 479 tasks, of which 19,152 come from its Tien Kung humanoid. They also include Pelican-VL 1.0 embodied-brain models (7B and 72B released 2025-11-13, with 3B and 235B-A22B checkpoints listed later), the XR-1 VLA (Apache licence, deployable on Tien Kung 1.0 and 2.0), and world-model, value-learning and data-augmentation sets released through 2026.

Unitree released the training code, inference code and weights of UnifoLM-WMA-0 on 2025-09-15, and deployment code for its robots on 2025-09-22. UnifoLM-WMA-0 is a world-model-action framework whose world model works both as a synthetic-data simulator and as a policy enhancer. Its LICENSE is CC BY-NC-SA 4.0. Unitree followed with UnifoLM-VLA-0, a VLA for humanoid manipulation (weights and code released 2026-01-29), demonstrated on Unitree G1 tasks. Its repository page shows no licence.

Fourier presented the N1 on 2025-04-14 as its first open-source humanoid. According to Yicai, the 1.3 m, 38 kg robot was released with a full bill of materials, structural blueprints, assembly guidelines, core control systems and basic operating software. Fourier said whole-body motion-control modules would follow.

Galbot is a co-affiliation, with Peking University, on GraspVLA, a grasping VLA pre-trained on synthetic data. The model and simulation playground were released on 2025-07-25, StereoVLA on 2025-12-29, and the SynGrasp-1B dataset (a billion frames, 240 object categories) in August 2026.

Alibaba's DAMO Academy released RynnVLA-001-7B, which learns manipulation from first-person human video, on 2025-08-08 under Apache-2.0. It released the RynnBrain family on 2026-02-02, also under Apache-2.0.

Several Chinese embodied models are built on Alibaba's open Qwen vision-language models rather than trained from scratch. RoboBrain 2.0 and Pelican-VL 1.0 both declare the qwen2_5_vl architecture on Hugging Face. RynnBrain lists Qwen3-VL-2B, 4B, 8B and 30B-A3B Instruct as its base models. Unitree's UnifoLM-VLA-0 README says much of its code is inherited from Qwen2.5-VL, Isaac-GR00T and other projects.

AgiBot's GO-1 uses InternVL2.5-2B from Shanghai AI Lab as its vision-language backbone, according to its technical report.

DeepSeek's role so far is commercial rather than on-robot. According to a stock-exchange filing reported by The Standard (2026-08-06), DeepSeek invested 140.8 million yuan in Unitree's Shanghai IPO, taking 933,399 shares or 2.31% of the strategic placement, and the two agreed to jointly develop AI models for humanoids. The same report notes that DeepSeek's flagship models are language-only and its vision-language research models are not integrated into them.

Licences split into two groups. The robot makers' own flagship releases (AgiBot's dataset, GO-1 and Genie Envisioner, and Unitree's UnifoLM-WMA-0) are under CC BY-NC-SA 4.0, which bars commercial use and requires derivatives to carry the same licence. Lab and cloud-company releases (BAAI's RoboBrain 2.0 and RoboOS, Alibaba's RynnVLA-001 and RynnBrain, X-Humanoid's XR-1 and RoboMIND) use Apache-2.0, which permits commercial use.

Several releases state no licence on the pages checked: UnifoLM-VLA-0, Pelican-VL, GraspVLA and the Fourier N1 package. Anyone planning to build a product on them should read each repository's LICENSE file first.