SenseTime Open-Sources Code2Video Model

SenseTime has released SenseNova 6.8 Code2Video, an open model that makes short videos by writing and running code instead of generating pixels directly. The 27-billion-parameter model is fine-tuned from Alibaba’s Qwen3.8-27B and has a 512K-token context window, according to its Hugging Face model card, where the weights are published under the Apache 2.0 licence.

In its announcement on X, SenseTime called it “a lightweight, open-source model that generates executable code to create and render short videos, with promising early results in visual composition and motion design”. The company has also published the prompts and full generation records behind its showcase videos, and says each one “was generated from a single prompt in one run”.

How code-driven video works

Most video models, such as Google DeepMind’s Genie 4, produce frames directly. Code2Video instead writes a program that draws each frame, renders it, looks at the result and revises the code, over many turns. In its GitHub repository, SenseTime argues that this fixes weaknesses of generative video. Every element is a named object that can be changed in a few lines, text and formulas are typeset exactly, and motion is computed from equations, so “the physics on screen holds up”. The model card says code can also call image, video, 3D and editing tools, so a program could chain them into one workflow.

SenseTime recommends running the model inside Anthropic’s Claude Code, using a published “code-video” skill that sets out a checkable workflow: write a specification, render every frame as a function of time, review contact sheets and verify the final file. The company says the model “sustains runs of more than 10 hours and more than 2,000 turns”. It suggests first expanding the user’s request with a stronger general model, saying “we use Claude Opus 5.5”, before Code2Video produces the film.

What is in the release

The release has three parts: the model weights on Hugging Face, the production skill and prompt templates on GitHub, and full production trajectories, the logs of messages, tool calls and review images from each showcase film. SenseTime gives examples for serving the model with LightLLM, vLLM or SGLang split across four GPUs. A 90-second launch film, and every video in it, was made by the model, the GitHub page says.

The model card carries a practical warning: “The run executes shell commands without asking, so start it in an isolated working environment.”

SenseTime is also candid about the limits. The model makes motion graphics, explainers and stylised 2D and 3D scenes, and “does not generate photorealistic footage”. There are no benchmark scores: “We report qualitative examples only,” the GitHub page says, adding that no systematic benchmark for code-driven video exists yet. SenseTime says the model “is clearly behind the strongest proprietary models on visual polish and long-horizon reliability”, and that “a production run still takes hours of agent time today”.

A small open model aimed at explainers, not footage

The release is narrower than a general video model, and that is its strength and its weakness. Code suits videos where exact text, numbers and timing matter, such as maths explainers, data stories and product animations, and where clients ask for small changes. Being able to edit one element without regenerating a whole clip is a real advantage over sampling-based tools, at least on SenseTime’s own showcase. The cost is time and compute: hours of agent work per film, a long context, and a 27B model that SenseTime’s own examples spread across four GPUs. Without benchmarks, buyers have only SenseTime’s chosen examples to judge from, so independent comparisons will matter. The choice of Qwen3.8-27B as a base also continues a pattern of developers building specialised models on Alibaba’s open weights, as Underdog did with its compressed Saluki 27B. For developers already using agents to turn screenshots into interface code, video is an obvious next target. It also adds to a run of open releases from Chinese teams this week, including DiffuSpace’s Clover diffusion language model.