SenseTime open-sources Code2Video, a 27B code-driven video model
The Apache 2.0 model writes, renders and reviews its own frames inside Claude Code, for motion graphics and explainers rather than photoreal footage.
Key takeaways
- SenseNova 6.8 Code2Video is a 27-billion-parameter model fine-tuned from Qwen3.8-27B, released under Apache 2.0 with a 512K-token context.
- It renders frames from code the model writes, inspects the result and revises, in runs SenseTime says can last more than 10 hours and 2,000 turns.
- SenseTime publishes no benchmark, reporting qualitative examples only and calling a benchmark for code-driven video an open problem.
SenseTime released SenseNova 6.8 Code2Video, a 27-billion-parameter open model that makes short videos by writing and running code instead of generating pixels.
The company announced it on X on October 28, 2026, and published the weights on Hugging Face and the production files on GitHub. Both the GitHub repository's first commit and the Hugging Face repository date to October 24, four days before the announcement.
Code2Video is fine-tuned from Qwen3.8-27B and carries a 512K-token context, English and Chinese support, and an Apache 2.0 license. It writes a program that draws each frame as a function of time, runs it, reviews the rendered frames as images, and revises the code until the film passes its own checks.
It targets motion graphics, explainers, stylized 2D and 3D scenes and procedural audio. SenseTime says the model does not generate photorealistic footage, so it is not a like-for-like swap for pixel video generators.
The release is more than weights. It ships deployment commands for LightLLM, vLLM and SGLang, each splitting the model across four GPUs, a code-video skill for Claude Code with a fixed workflow and eight craft guides, a prompt-enhancement template, and the full production trajectories behind its showcase films.
SenseTime describes three stages. A strong general model expands a short request into a dense brief, which in SenseTime's own tests is Claude Opus 5.5. Code2Video then produces the film inside Claude Code using the code-video skill, and the creator can ask for changes in the same project. Each run delivers final.mp4, SPEC.md and VERIFY.md, and SenseTime says a run can last more than 10 hours and more than 2,000 turns.
The documentation lists its limits. SenseTime reports only qualitative examples, says it has not built a systematic benchmark for code-driven video, and calls one an important open problem. It also says the model is clearly behind the strongest proprietary models on visual polish and long-horizon reliability.
Independent write-ups draw the same line. Neural Brew reports there is no independent evaluation yet, and llms.blog puts Code2Video in the territory of tools where exact text and repeatable edits matter more than photorealism.
That makes this a different kind of release from a pixel video model. Its case is that a film built from code is easier to edit and check than a sampled clip, because a change touches a few lines instead of the whole generation. It also runs inside Claude Code, so part of the pipeline leans on another company's proprietary model.
SenseTime says it will scale the model and extend it to editing, compositing, animation and 3D.
The weights are free under Apache 2.0 at huggingface.co/sensenova/SenseNova-6.8-Code2Video. The code-video skill installs into .claude/skills/code-video in a task folder; the prompts and trajectories are in the OpenSenseNova GitHub repo.
Sources
- huggingface.co - model card: parameters, base model, license, three-stage pipeline, limitations
- github.com - repo: README, code-video skill, eight craft guides, run length, limitations
- x.com - SenseTime's own announcement, dated October 28, 2026
- sensetime.com - company profile: founded 2014, listed on the Hong Kong exchange
- neuralbrew.ai - independent write-up: no independent evaluation yet
- llms.blog - independent write-up on who the tool suits