36Kr reports on a 4B-parameter "world model" from Kaiwu (开悟) that sweeps four embodied AI benchmarks, redefining SOTA. The standout: a 4B model matches or beats 70B+ models on embodied tasks, marking a major milestone for "small specialist" embodied AI.
The "4B sweep four benchmarks" highlight: the Kaiwu world model, with only 4B parameters, hits SOTA on four major embodied AI benchmarks — RoboNet, Habitat-Sim, AI2-THOR, and SAPIEN. The previous SOTA on these benchmarks was held by 70B+ general-purpose models. The 4B model is competitive or better on all four, with 15-20× lower inference cost.
The "world model" architecture: the model is a "video world model" — it takes a video of the current state and predicts the next state. The model is trained on a large corpus of robot manipulation videos, with a focus on "long-horizon consistency" (the model can predict 30+ seconds of future state with high accuracy).
The "small specialist" insight: the 4B model is significantly more efficient than 70B+ general models because it's "specialized" — it doesn't waste capacity on general capabilities (chat, code, reasoning), and all 4B parameters are dedicated to the embodied task. The "specialist" approach is significantly better than "generalist" for domain-specific tasks.
The benchmark: on the "long-horizon manipulation" benchmark, the 4B model hits 84.2, on par with the 70B SOTA (85.1). On "object navigation," it hits 78.5, on par with the SOTA (79.2). The inference cost is 15× lower, making the 4B model significantly more practical for production deployment.
The bigger takeaway: "small specialist world models" are the right architecture for embodied AI. The "bigger is better" assumption is breaking, and the next round of embodied AI competition will be in "small specialist" quality. For the industry, this signals that "world model" vendors should focus on specialist models, not general-purpose ones.