Meituan releases LongCat-2.5, a multimodal model with 1.6 trillion parameters for coding long-horizon agents

Dawa Technology, a company based in Chongqing that originated from the field of virtual film production, has introduced Scenes AI—a real-time world model. This system is capable of transforming video shot with a regular camera into an editable and persistent world.

The product received an excellence award at NVIDIA's intermediate startup event in Suzhou on October 15. Thirty members of NVIDIA Inception participated in this stage among more than 580 attendees.

Most video models focus on predicting the next frame. However, Scenes AI supports a continuous world state. Elements such as the scene, people, camera, lighting, objects, and physics are stored as separate variables that can be read and modified independently. When a change occurs, only the corresponding part of the state is updated, not the entire image redrawn.

In demonstrations, Dawa showed how a room transforms into a hand-drawn meadow while the chair, teddy bear, and wardrobe maintain their positions. Wall inscriptions remain in place, and people appear and disappear, while the weather changes from snow to rain.

The system includes a complex control layer and a lightweight generative layer. The control layer, built around the Latent Stage module, separates the scene, character, camera, lighting, costume, and props in latent space, applying physical constraints for solids, fabrics, liquids, and particles. The rendering layer, LiveStage, uses causal sparse diffusion to sample only the changing areas.

The model has a total of 34 billion parameters but activates about 4 billion per step, which is approximately 11.8 percent. Dawa claims this reduces GPU usage by 40–60 percent and accelerates inference three to five times compared to traditional video diffusion.

On a single consumer NVIDIA GeForce RTX card, the system achieves real-time output at 720p resolution and 25 frames per second. In continuous editing tests, the company reports a 98.2 percent success rate for target editing and a 93.3 percent preservation of human identity. The system utilizes CUDA, TensorRT, and NVENC and operates on local hardware without cloud computing.

Dawa states that its data was gathered from an 84,000 square meter studio space, over 300 real productions, and 180 TB of world state data. The company's Chief Product Officer, Li Zongxin, summarized the approach as follows: 'Video can end, but the world must keep running.' All presented performance metrics are data provided by the company itself.

The global models team at ShengShu Technology has started a live stream of the entire training process for Motubrain 2.0, the next version of its embodied artificial intelligence model for robots. The broadcast began on October 21st and is scheduled to last about two weeks.

The company announced that Motubrain 2.0 will soon be released to the public and decided to demonstrate the training process before this release.

The Motubrain models represent a family of fundamental robot models from ShengShu, aimed at providing robots with a universal brain capable of applying skills across various tasks, scenarios, and robot types. The first version of Motubrain was introduced in April as a global model. It is worth noting that ShengShu is more widely known for its Vidu video generation models.

Stream Structure

The live stream is divided into four stages. The pre-training stage explores how training scale affects performance and zero-shot generalization ability. Subsequent training focuses on performing specific tasks to improve execution. This is followed by an agent task stage, which tests whether the model can transfer acquired knowledge to other assignments. The final stage is dedicated to recursive self-improvement to determine if the model can evaluate its actions based on execution feedback and improve them.

The real-time streaming page displays the training steps and success metrics as they are updated.

The main idea behind this demonstration is transparency. Progress in embodied AI is usually assessed by external observers through polished robot demonstrations, benchmark tables, and scientific papers, which rarely reveal exactly how capabilities are formed during training or what problems arise. ShengShu claims that the stream will show not only successes but also errors, the trial-and-error process, and adjustments, and hopes to share open questions regarding scaling, cross-task generalization, and autonomous improvement with other researchers.

This step aligns with a broader trend among Chinese AI teams who publicly share their training progress. Previously, Luo Fuli, head of Xiaomi's MiMo models, published details of the reinforcement learning training phase for MiMo-V2.6.

Nevertheless, there are limitations to what a live stream can prove. The model weights have not yet been released, and the company itself states that the research remains exploratory, and the results will require subsequent experiments and evaluation for confirmation. The announcement was made by ShengShu and covered by Chinese tech media. The true test will be the public release and whether its benchmarks match what was shown in the stream.

VAST, a company specializing in generative AI in 3D and behind the Tripo platform, has made the EdenLive model publicly available. This model operates in real-time and is capable of taking simplified game frames or existing video, continuously re-rendering them into detailed images in a chosen style, which can range from photorealistic to anime.

The release includes 5B and 14B model weights, code for real-time inference, a technical report, and end-to-end usage demonstrations. It is important to note that EdenLive does not simulate the world on its own; instead, a higher-level program, such as a game engine or code written by a coding agent, is responsible for enforcing rules, tracking world state, and outputting a rough reference video.

After receiving this reference video, EdenLive performs the rendering, guided by the initial frame and a text prompt. VAST asserts that this separation allows game logic to remain explicit and testable, while concentrating the model's limited real-time capabilities solely on the visual appearance.

The approach was first presented in June as part of a larger world research effort known as Project Eden. According to VAST, the 5B model can operate at a frame rate above 24 frames per second at 720p resolution on a single Nvidia RTX 5090 GPU. The README file on GitHub specifies this figure as approximately 24.7 FPS at a resolution of 1280 by 704 pixels, utilizing acceleration from the LightX2V inference framework.

The technical report reports a rate of 29.8 FPS at the same resolution on a single H100, as well as 26.8 FPS at a resolution of 832 by 480 for the 14B model running on eight H100s. VAST states that the total latency from user input to the corresponding frame is in the hundreds of milliseconds, and the system can generate content continuously for several minutes.

Both models are based on the Wan2.2-TI2V-5B and Wan2.1-I2V-14B video models. The inference code and core diffusion, VAE, and decoder weights are distributed under the Apache 2.0 license, although not all components have this license. The visual encoder uses the DINOv3 license, and the image model used in the request-to-game workspace is under the Non-Commercial FLUX License, which restricts the use of this workspace to non-commercial purposes.

VAST also introduced Game Prototype Turbo—a skill that transforms a text prompt into a rough playable game, which EdenLive can then render. VAST demonstrations include a platformer where a character freezes lava to create waypoints, a two-player fighting game where health bars and round timers are tracked by a higher-level system, and video rendering that transforms a person filmed against a blue background into a cartoon character. Furthermore, the model demonstrates replacing human hands with robotic ones in first-person footage to create additional training data for embodied AI.

VAST notes that EdenLive is currently best suited for low-frequency interactions, such as scene inspection and box-pushing puzzles. Fast, intense gameplay, maintaining multi-character consistency, fine control over small objects, and visual stability during long sessions still require refinement.