Unitree Robotics has begun publishing partial weights and corresponding datasets for UnifoLM-WLA-1.0. This model is a foundational model for humanoid robots, possessing approximately 6 billion parameters and designed to handle language and real-world actions. Previously, the company presented the model architecture and an evaluation matrix across 64 tasks.
According to a report from a Chinese tech publication on September 20th, some components of this system are now available for download on the project page. However, the trained code, the UnifoLM-WLA-Base weights, the full-body teleoperation corpus, and broader operational datasets are still pending.
This phased open release updates a previous announcement on the project page where the Code, Models, and Datasets were initially marked as 'coming soon.' Furthermore, the software aspect of the development remains separate from Unitree's recent G1+ hardware update.
One control sample was trained on approximately 2500 hours of data gathered from a real robot using Unitree's open datasets and publicly available sources such as BitRobot HIW-500. It is designed to cover 10 whole-body tasks plus 54 tabletop tasks, without requiring a separate policy launch for each routine operation.
Examples of household tasks provided by Unitree include folding towels, arranging plates, charging batteries, washing clothes, and cleaning sofas—tasks that require synchronization between hands, end-effectors, and legs. A unified action space utilizes end-effector pose tokenization, hand joint, and lower body joint tokenization via residual vector quantization, allowing these streams to remain temporally consistent.
The presented tests support parallel two-finger grasps and multiple five-fingered hands. The MMDiT action expert is the UnifoLM-ER-1-4B model, which is an embodied reasoning model. It is built upon Qwen3-VL-4B and incorporates over 5 million spatial samples covering object detection, detection, multi-image reasoning, and trajectory prediction.
Unitree reports that its model leads among open models in seven out of sixteen multimodal perception benchmarks compiled by the company, including Where2Place, EmbSpatial, and RoboSpatial. Some metrics are comparable to those of proprietary systems in the same tables. It should be noted that these figures are presented by the company itself and have not yet been reproduced by a third party, nor do they inherently prove success in real homes.
For integrators, the practical takeaway is the phased release: partial weights are currently available, while the complete code and data corpora are still awaited. Until the remaining resources arrive, any claims about task transfer to non-Unitree robots should be viewed as a roadmap rather than a ready-made third-party kit.
The Shanghai Artificial Intelligence Laboratory has released the Intern W0 physical world model. This stacked system is oriented towards robotics and is designed to transfer the behavior of large models from digital planning to physical, contact-rich operation.
The laboratory emphasizes that W0 was initially developed with the goal of integrating data from vision, force, and touch. This allows the robot to perceive, predict, act, and correct its actions in parallel, instead of following a rigid cycle: first perception, then planning, then movement.
This model differs from the laboratory's previously published collaboration with LUMIA Lab from Shanghai Jiao Tong University, which focuses on predicting next-level concepts rather than force-based physical control.
The release of W0 is defined by two key elements. The first is native sensory perception of force and touch, which integrates contact feedback into the action selection and perception process alongside data from vision and body proprioceptive state. Vision determines approach directions and target points, while force and touch update grip quality and register slippage after contact begins. Proprioceptive state tracks position and movement.
The second principle is duplex interaction, which maintains a continuous flow of multimodal inputs, including language instructions, visual observations, force/touch feedback, and body state. Simultaneously, the model generates semantic understanding, predictions of future states, and motion commands. The asynchronous architecture with a multi-speed mode allows for background updating of long-term planning while high-frequency action updates continue to react to fresh observations, thereby reducing the interval between environmental change and motion correction when vision alone is insufficient.
Researchers note that this combination helps in tasks requiring high precision, such as determining grip stability, correctness of contact settling, or how the next micro-correction should look. Model W0 is already integrated with the Intern InkStone scientific discovery platform—the English product name presented on the lab's portal. It functions jointly with the Intern S2 large science model.
Together, they have provided support for closed loops in both wet and dry laboratory conditions. These loops covered directed evolution of proteins for gene editing, organic synthesis of mepivacaine, and preparation of lipid nanoparticles, including dynamic tuning of process parameters during nanoparticle work, which is mentioned in secondary sources as early evidence of science-oriented physical AI.
For groups involved in chemistry, biology, and materials science, W0 is viewed as a tool for automated experimental cycles, not as a consumer robot demonstration. It remains to be seen how broadly the duplex force and touch control generalizes across different laboratory instruments and wet lab protocols, which will determine whether the Intern stacked set becomes the default layer in Intern InkStone workflows or remains tied to previously demonstrated closed-loop cases. The laboratory presents these cases as the beginning of the path from schematic derivation to instrumental experimentation, not as a ready-made product catalog.
On September 10th, Unitree introduced UnifoLM-WLA-1.0, which is a generative foundational model for next-generation humanoid robots with approximately 6 billion parameters. According to the company's statement, a single control file trained on data from a real robot spanning about 2500 hours can coordinate 64 tasks, covering 10 whole-body manipulations and 54 tabletop manipulations, as well as operate with both two-fingered grippers and multiple dexterous five-fingered hands.
This release differs from the previous demonstration of Unitree UnifoLM-X2 sparring; UnifoLM-WLA-1.0 represents a broader 'vision-language-action' stack for performing daily household and tabletop tasks, rather than a combat-oriented demonstration.
Architecturally, UnifoLM-WLA-1.0 is based on interaction-centric world modeling augmented by an MMDiT action expert. This embodied reasoning stage, named UnifoLM-ER-1, is built upon Qwen3-VL-4B and was trained on over 5 million embodied samples. These samples cover aspects such as point prediction, object detection, multi-image reasoning, 2D trajectories, 3D detection, and answering spatial questions, enhancing spatial perception while jointly training with general image and text data to maintain broad 'vision-language' skills.
Furthermore, dynamic region prediction utilizes optical flow to extract future scene changes, compresses them using VQ-VAE into fixed-length mask tokens, and trains the 'vision-language' model to predict which regions will start moving after interaction begins. Discrete action learning via residual vector quantization segments the unified action space into end-effector poses, arm joints, and lower body joints, aligns these tokens across common time steps, and feeds them into UnifoLM-ER-Flow. The WLA model they released includes the MMDiT expert for decoding continuous robot actions from this multimodal framework.
In spatial and embodied reasoning tests conducted on 16 multimodal datasets, Unitree reports that UnifoLM-ER-1-4B outperforms open models in seven benchmarks. Additionally, on overlapping spatial datasets, it surpasses the locally proprietary GPT-6 Astra in metrics such as Where2Place, BLINK spatial subtasks, CV-Bench, and EmbSpatial. Training relies on Unitree Open Datasets and BitRobot-HIW-500, as well as other real robot data sources. Video evaluations are demonstrated on Unitree G1 hardware performing tasks like folding towels, packing phones, making beds, sorting shoes, and cleaning bathrooms—tasks requiring coordinated movement of arms, end-effectors, and lower body within a single policy.
It is important to note that the project page is available to developers, but the code, models, and datasets are currently marked as 'Coming Soon.' Unitree has announced its intention to make the project open source and published documentation on the architecture and task matrix, but the weights and data corpora are not yet available for download. Until these resources become available, the practical signal for researchers and integrators is the disclosed roadmap of the general humanoid foundation and the benchmark table, rather than a ready-to-run control file for third-party robots.
At the LEAP 2026 event on September 3rd, HUMAIN, supported by the Public Investment Fund (PIF) of Saudi Arabia, unveiled the HUMAIN-M3 model. This model is focused on the Arabic language and is currently available for preview on the HUMAIN Node platform, with plans for an open-weights release later.
The foundation for HUMAIN-M3 is the open mixture-of-experts (MoE) structure M3 from the Chinese startup MiniMax, which was released approximately three months prior. The HUMAIN-M3 model retains the architectural features of M3, possessing a total of about 428 billion parameters, with approximately 23 billion parameters active per token. Furthermore, the model undergoes additional training on over one trillion native Arabic tokens.
According to seven public Arabic benchmarks published by HUMAIN, this model achieved five top rankings and an average weighted score of 89.37%. This result is 9.03 points higher than that of the base M3 model and surpasses the claimed results of GPT-5.6 SOL and Claude Opus 5. The main emphasis is on speed: it utilizes general reasoning, multimodality, agents, and long context from the open base model, followed by local language and cultural specialization.
This approach aligns with the broader global race for sovereign artificial intelligence. According to data from CNAS, tracked in Chinese publications, 184 state-sponsored sovereign AI projects were registered in 67 countries by mid-2026, with 41 projects added in the first half of the year. The number of open bases for local continued training has more than doubled since the end of 2024. Buyers are focusing on dialect mixing, on-premise data storage, and the risk of remote API disappearance following export control shocks, which favors downloadable weights and local post-training.
MiniMax's international activities also follow this trend. The company states that its products are already available in over 230 countries and regions. In Europe, MiniMax collaborates with SambaNova for high-speed inference on SN50 hardware, as well as with Nebius Token Factory as a specialized hosted open solution. HUMAIN pushes this concept further: from simple downloading and APIs to a national model program that invests proprietary language data and computational power on top of the Chinese benchmark.
For laboratories working on open models in China, developers participating in rankings are no longer the only clients. Governments now demand a starting benchmark model that they can own, update, and run independently. The integration of MiniMax M3 into HUMAIN-M3 is an early and concrete example signaling that open weight delivery, hardware adaptation, and long-term support can be as important as the next benchmark score. The race outside the US and China duopoly is increasingly about who provides the reusable base, rather than who wins each English language test.
