Skywork AI Launches Three Major AIGC Models to Tackle Core Industry Challenges
On March 27, 2026, at the 2026 Zhongguancun Forum held in Beijing, Skywork AI (a subsidiary of Kunlun Tech) unveiled a suite of three generative AI models targeting different modalities: the game world model Matrix-Game 3.0, the large video model SkyReels V4, and the large music model Mureka V9. This initiative aims to systematically address key technical challenges in AIGC for gaming, video, and music, including long-sequence consistency, audio-video synchronization, and content controllability.
Matrix-Game 3.0: An Interactive World Model Achieving Long-Sequence Consistency
World models represent a frontier in the AI field, aiming to enable AI to understand and simulate the laws of the physical world. However, existing models often suffer from “memory loss,” failing to maintain logical consistency in scenes after prolonged interaction. Matrix-Game 3.0 is designed to address this core issue.
To overcome data bottlenecks, Matrix-Game 3.0 employs a dual-pipeline data engine: one pipeline uses Unreal Engine 5 to automatically generate high-fidelity interactive data, including camera poses and action commands; the other automatically extracts dynamic interaction data from AAA games, providing vast, high-quality material for model training.
At the architectural level, the model introduces three key technologies:
- Memory Injection: When generating a new frame, the model retrieves and references historical “memory frames” to ensure that the environment remains unchanged when a user returns to a previously visited location in the scene.
- Error Buffer Anti-Drift Mechanism: By explicitly modeling and injecting errors during training, the model learns to self-correct during long-sequence generation, preventing state deviation.
- Multi-Step Autoregressive Distillation (Distillation Acceleration): This technique significantly reduces the number of inference steps, allowing a lightweight 5B parameter model to achieve real-time interactive generation at 40 FPS at 720p resolution, while maintaining minute-level temporal consistency.
Furthermore, a larger MoE-28B model further enhances generalization capabilities. The open-source release of Matrix-Game 3.0 positions it as a competitive solution alongside closed-source models like Google DeepMind’s Genie 3.
SkyReels V4: Solving Audio-Video Synchronization and Content Control from the Ground Up
Traditional AI video generation pipelines often process visuals and audio separately, leading to issues like poor synchronization and mismatched lip movements. SkyReels V4 fundamentally solves this problem with its proprietary Dual-Stream MMDiT architecture. This architecture uses a bidirectional cross-attention mechanism to co-generate video and audio within a single model, ensuring high synchronization from the very first frame.
For content controllability, SkyReels V4 integrates multimodal reinforcement learning and a precise control system. Users can provide keyframes or image grids (up to 9 images) as references to guide the model in generating videos with coherent plots and consistent characters. The model also supports automatic camera switching and dialogue matching for multi-character conversational scenes, covering multiple languages including Chinese, English, French, and Japanese. Additionally, it features practical post-editing functions such as watermark removal, object removal, and character swapping.
Mureka V9: Enhancing the Precision and Professionalism of AI Music
Building on the MusiCoT (Music Chain-of-Thought) technology from its predecessor, Mureka V8, Mureka V9 shifts its development focus from “generation” to “control.” The new version offers more precise paragraph-level lyrical semantic control, ensuring correct vocal-lyric matching. It also enhances the professional quality of the audio mix and reduces unnecessary humming, resulting in more natural musical pauses and rhythms.
According to subjective evaluation data, Mureka V9 surpasses competitors like Suno V5 across multiple dimensions, including musicality, expressiveness, and arrangement. Its score for “precise control” significantly increased from 6.93 to 7.24, validating the effectiveness of its technological upgrades. At the product level, Mureka supports a version-based iterative workflow, allowing creators to make partial modifications to elements like melody and vocals, which greatly enhances creative efficiency and freedom.
The 2026 AGI Strategy: Building an Ecosystem from Models to Platform
At the event, Skywork AI’s Chairman and CEO, Zhou Yahui, announced the company’s “2026 AGI Strategy,” centered on a “3+1” framework. The three major models (Matrix-Game, SkyReels, Mureka) will serve as technology engines, empowering the gaming, video, and music industries, respectively. The “+1” refers to the Skywork Super Agent, which will act as a unified platform integrating all AIGC capabilities. It will offer a one-stop creation tool for consumers (B2C) and open its API to developers (B2B).
This strategy aims to build a full-chain ecosystem—from content generation and editing to distribution—through a product matrix including the AI short-drama platform Dramawave, the AI music platform Mureka, and the AI gaming platform Maosen Academy 2.0. The goal is to usher in the “era of one-person companies” proposed by Zhou Yahui. The company’s roadmap indicates 2026 as a year for infrastructure development, 2027 for fostering ecosystem growth, and 2028 for establishing a mature platform economy.
