SenseNova 6.8 Code2Video: SenseTime releases under Apache 2.0 a 27-billion-parameter model that makes videos by writing code
SenseTime has released SenseNova 6.8 Code2Video, an open 27-billion-parameter model derived from Qwen3.8-27B. It does not generate pixels: it writes and fixes the program that computes each frame. The developer publishes no benchmarks and acknowledges a clear gap behind the best proprietary models.
SenseTime officially presented on X this Wednesday, October 28 at 3:09 PM Paris time, SenseNova 6.8 Code2Video, whose weights had been uploaded the day before to Hugging Face. The Chinese model takes the opposite approach from current video generators: instead of producing pixels, it writes the program that draws each frame, runs it, looks at the result, and corrects it.
The model has 27 billion parameters, is fine-tuned from Qwen3.8-27B, accepts a 512,000-token context, and is released under the Apache 2.0 license, with no commercial use restrictions.
A video you edit like a program
In a conventional generator, the video comes from a single random draw. Changing the color of a title or the rhythm of a transition forces you to rerun the generation, and the rest of the shot changes with it. SenseTime lists three criticisms in its documentation: plausible but physically wrong motion, videos that are hard to retouch, and on-screen text, formulas, or numbers that are often distorted.
With Code2Video, each element is a named object in the code and each frame is a function of time: the same code always renders the same film. A tweak (“switch the set to pastel tones,” “add a plane”) becomes a change of a few lines, and everything unaffected stays identical. The repository thus shows an airport slat panel reworked in two rounds of notes.
The targeted uses are specific: motion graphics, explainer videos (math, data), stylized 2D and 3D scenes, animated typography, small characters, procedural audio synced to music. SenseTime states it plainly: the model does not generate photorealistic images.
Built to run for hours in Claude Code
Code2Video is not a one-click tool. The described pipeline has three steps: a powerful general-purpose model first turns a short request into a detailed brief (SenseTime uses Claude Opus 5.5), then Code2Video produces the film in Claude Code, Anthropic’s coding agent, with a provided code-video “skill” (eight production guides, rendering and verification tools), and finally the creator can request corrections in the same project.
The developer announces sessions of more than 10 hours and more than 2,000 turns, and provides deployment commands with vLLM, SGLang, and LightLLM spread across four GPUs. It also publishes the complete production trajectories (messages, tool calls, control images), useful for understanding or reproducing the method. You need Node.js, Python, ffmpeg, and a Chromium browser driven by Playwright for rendering.
What SenseTime has yet to prove
The documentation is candid about its limits. There is no published benchmark: only selected examples, including a 90-second launch film presented as produced by the model. SenseTime acknowledges a clear gap behind the best proprietary models in visual polish and reliability on long tasks, and notes that a production still takes hours of agent work.
Our take
Code2Video does not compete with Kling, Seedance, or Veo: it cannot film a realistic scene. Its value lies elsewhere, in the territory of tools like Manim or Remotion, where text, a formula, or a chart must be exact and where you want to retouch a detail without regenerating everything. It is also a serious lead for creators of educational videos and motion design who want a reusable source file rather than a frozen render. The barrier to entry remains: four GPUs, hours of compute per film, and a proprietary model upstream to write the brief. Without independent evaluation, we will judge on the merits, with the published trajectories as the first way to verify what the model actually does.
Sources
- SenseTime (X) — SenseNova 6.8 Code2Video announcement Primary source
- Hugging Face — model card sensenova/SenseNova-6.8-Code2Video Primary source
- GitHub — OpenSenseNova/SenseNova-6.8-Code2Video repository Primary source
A primary source comes directly from the company, lab or organization concerned. A secondary source reports the information without being its author: we flag it as such. Spotted an error? See our corrections policy.
