SenseNova 6.8 Code2Video

SenseNova 6.8 Code2Video is a lightweight open model from SenseTime for code-driven video creation. Instead of sampling a clip from a video diffusion model, it writes executable code that renders each frame, in the manner of Manim-style motion graphics. It is fine-tuned from Qwen3.8-27B and has a 512K context window. The current release focuses on short videos implemented entirely in code.

Because every element is a named object in the code, individual elements can be resized, retimed or restyled without changing the rest of the film. After a first cut, the model accepts follow-up notes, edits the same project and re-renders. Text, formulas and data are typeset exactly, and each frame is a pure function of time, so the same code reproduces the same video. The authors say rendering on ordinary hardware is cheaper than sampling a video model, though a full production run still takes hours of agent time.

The GitHub repository releases model weights (hosted on Hugging Face), deployment and serving instructions, a code-video skill with workflow guides and rendering/checking tools, a prompt-enhancement template, and full production trajectories. It is aimed at researchers and developers. SenseTime describes it as a first step; long-term plans include calling image, video and 3D models from code and extending to editing, compositing and animation. The license is not stated in the provided material.