Fresh neural networks of 2026: an overview of what has just been released

Seedance 2.0
Seedance 2.0 is a new video generation model from ByteDance, officially launched in March 2026. It accepts text, images, audio, and video simultaneously, generating cinematic clips with native sound, realistic physics, and camera control. By the end of March 2026, it ranked first on the Artificial Analysis leaderboard with an Elo score of 1,269, surpassing Google Veo 3 and Runway Gen-4.5.

Kling 3.0
Kling 3.0 from Kuaishou was released with significant improvements: photorealistic rendering, videos up to 15 seconds long, and native sound generation in multiple languages and dialects. It is particularly praised for natural human movement, lip synchronization, and accurate facial expressions—the best results among all current video models. It is already available via a global API at a price of approximately $0.075 per second.

MiniMax M2.7 / Hailuo AI
In February 2026, MiniMax released the M2.5 model, followed by M2.7 in March. The flagship Hailuo 02 leads in portrait animation: it transforms static photos into smooth videos with effects like wind in the hair, fabric movement, and natural facial expressions in 1080p resolution. It works without a VPN in Russia and understands Russian prompts.

Vidu 2
Vidu 2 is a powerful all-in-one video generator from 2026 by Shengshu AI from China, specializing in text-to-video, image-to-video, and reference-to-video with good physics, consistency, and fast rendering. The model uses a "smart cuts" system—the AI intuitively understands the narrative and automatically changes the camera angle if a character is running or if the scene requires a change in perspective. This is why Vidu is called the most "artistic" of the new video models.

GLM-5.1
Z.ai quietly released GLM-5.1 on March 27, 2026, and its benchmark numbers are noteworthy. The model is designed for "agentic intelligence"—multi-step reasoning, planning, and coding—and is fully trained on Huawei Ascend chips without using American GPUs. This makes it an important player in the context of sanctions on American chips.

Qwen 3.5
The new series from Alibaba supports 201 languages and dialects—twice as many as the previous version. The open-weight version contains 397 billion parameters, has native agentic capabilities, and according to claimed benchmarks, is on par with GPT-5.2 and Claude 4.5. It was released in February 2026 and is available for free as open-source.