The Beijing research institute Beijing Academy of Artificial Intelligence (BAAI) has introduced a novel world model called Orca, which derives text, images, and robot commands from a single internal representation. The system was trained on 125,000 hours of unlabeled videos and 160 million event descriptions. In benchmarks, the compact four-billion-parameter version significantly outperforms much larger competing models.
Two learning methods create a common world state
Unlike classical AI models, Orca does not predict individual tokens, pixels, or robot signals, but learns an abstract internal state of the world – comparable to an internal representation of how an observed scene will change next. The team combines two learning methods for this purpose: In unconscious learning, the model observes dense, natural state transitions from continuous video sequences without them being described linguistically. In conscious learning, on the other hand, Orca processes 160 million event descriptions that link individual video segments with short text instructions. A frozen pre-trained language model called Qwen3.5 serves as the basis, on whose representations three training objectives act simultaneously: the prediction of pure observation transitions, the linguistically conditioned state prediction, and classical question-answer tasks. Only after that does the team attach lightweight, task-specific decoders to the frozen model, which generate text, images, or robot commands from the common representation. According to BAAI, a simple relationship applies: the stronger the learned internal representation, the better the downstream results are in all three output types.
Orca outperforms larger models in text and image benchmarks
In four language benchmarks for video understanding – including MVBench and TemporalBench – the compact four-billion-parameter version of Orca achieves an average of 51.8 percent, placing it ahead of the similarly sized model Qwen3.5-4B with 46.7 percent. The gap is even more pronounced compared to the more than eight times larger model Emu3.5, which achieves only 29.8 percent on average, despite having significantly more parameters. The lead is particularly large in tasks related to state transitions and motion sequences in videos, areas where an internal world model is said to be particularly helpful according to BAAI. Orca also performs well in image prediction: combined with an additional two-billion-parameter decoder, the system achieves a score of 59.8 points on the test set PRICE-V0.1, placing it ahead of the specialized image generator FLUX.2 small with 56.1 points. The generated images were evaluated by several other AI models as reviewers, including Gemini 3.1 Pro and GPT-5.4. These results come from the technical report of the BAAI team and are independently unverified.
Few examples suffice for robot control
The model’s ambition is most clearly demonstrated in robot control. For five real manipulation tasks, the team needed only 200 training examples per task to adapt Orca to a robotic arm. This allowed the system to reach a performance level comparable to the specialized robotics model π0.5. In error detection and correction during object manipulation, Orca reportedly even surpassed the comparison model according to BAAI. Notably, no single action-labeled recording was used during the actual pre-training; the model learned motion sequences solely from unlabeled videos. Only in the final adaptation phase were the few action-labeled examples added to map the learned internal representation to concrete control commands for the robotic arm. According to the team, performance grows continuously across all three task types with model size and the amount of video data: between 0.8 and 4 billion parameters and between 12,500 and 125,000 hours of video material, the training loss steadily decreases, with no upper limit yet in sight.
BAAI president sees world models as the next step in robotics
BAAI president Wang Zhongyuan explained the strategic direction of the institute even before the paper’s publication in an interview with the business medium 36Kr. Vision-language-action systems that translate camera images directly into robot control commands are useful for individual scenarios but reach limits in long task chains and complex spatial reasoning. Therefore, according to Wang, world models that represent cause and effect as well as temporal consistency are the next developmental step in robotics, far beyond individual control commands. Orca thus joins a growing number of Chinese world model projects seeking direct comparison with U.S. systems, after BAAI had already made waves with the openly available robotics model RoboBrain. BAAI has not yet released code or model weights for Orca; currently, only the technical report and a project page with example videos of the robot trials are available. This means that the work cannot currently be replicated by independent research groups or tested in their own applications, making direct comparison with Western world model projects difficult for the time being.
It will be crucial whether the results hold up outside controlled benchmarks, such as with objects or environments not present in the training material. It also remains to be seen whether BAAI will release code and weights, as only the technical report is currently available, without which independent verification of the numbers is hardly possible.
