SenseTime open-sources SenseNova 6.8 Code2Video

SenseTime open-sourced SenseNova 6.8 Code2Video, a lightweight model that turns a written description into a short video by generating executable code that then renders the result. Users describe an idea, the model writes code that builds and renders the video, and they can refine it with follow-up requests or feed in a reference video to recreate and remix in editable code form. SenseTime says early results show promising visual composition and motion design, and the company open-sourced the model weights, an accompanying "Skill" workflow, and the prompts and full generation logs behind its showcase videos.

The model is fine-tuned from Qwen3.8-27B. SenseTime frames code as a bridge between an idea and a finished creative work, since code can also call out to separate image, video, and 3D generation tools, chaining them into one production pipeline. The company says future research will extend the approach to video editing, animation, and 3D content creation.

SenseTime recommends running the model inside Claude Code as the agent harness, noting that producing a complete film is a long-running task: the model can sustain runs of more than 10 hours and more than 2,000 turns as it writes, renders, reviews, and revises code for a video project.

Why it matters

Treating video generation as a code-writing problem, rather than pure pixel generation, lets a model reuse existing software tools and plug into agent coding harnesses.

Source: SenseTime on X