SenseTime open-sourced a video model whose output is code

SenseTime open-sourced SenseNova 6.8 Code2Video on October 28. It has 27B parameters, an Apache 2.0 license, and a 512K context; the weights, configuration, and tokenizer are all on HuggingFace.

It takes a different path from common video models. Diffusion models generate pixels directly; it generates code, and after the code runs, it draws each frame of the video. According to the official model card, the model builds and renders videos by generating executable code, and it currently focuses on short videos implemented entirely in code.

The most direct impact of this difference is controllability. The original post gives an example: you say "change the title to blue," and only those lines of code change; the rest of the video stays untouched. You don't have to rerun the entire render twenty times for one detail and gamble that nothing else breaks this time.

In the traditional workflow, changing a frame is like pulling from a gacha. Code2Video turns video into a code project; changing one place is a diff.

It is suitable for motion graphics, explainer videos, charts, and animated text, but not for real footage. In other words, it doesn't handle live-action material; it codes ideas into videos.

The model is fine-tuned from Qwen3.8-27B. The HuggingFace page recommends Claude Code as the runtime environment, because a complete film is a long-horizon task: write code, render, inspect frames, revise, round after round. The official configuration is a 512K context, an automatic compaction threshold of 0.8, and 64K per output; it can run for more than 10 hours and more than 2,000 rounds at a time.

{
  "env": {
    "CLAUDE_CODE_MAX_CONTEXT_TOKENS": "524288",
    "CLAUDE_AUTOCOMPACT_PCT_OVERRIDE": "80",
    "CLAUDE_CODE_MAX_OUTPUT_TOKENS": "64000"
  }
}

The production workflow has three stages. In the first stage, a strong model is used for prompt enhancement; they use Claude Opus 5.5. In the second stage, the enhanced prompt is passed to Code2Video, which calls the code-video skill inside Claude Code to produce final.mp4, SPEC.md, and VERIFY.md. The third stage is optional: a human writes notes, and the model revises a new version within the same project.

To study how it works, trajectories/all.openai.jsonl stores one complete record per film per line, including reasoning, tool calls, results, retries, and compaction summaries. The 357 process images are in trajectories/figures/. GitHub has the production skill, including the workflow, craft guide, and rendering tools.

For deployment, it provides three examples: LightLLM, vLLM, and SGLang, all using 4-GPU tensor parallelism. The barrier to entry is not low; ordinary users may have to wait for a hosted service.

Using code as the intermediate layer for video has a direct benefit: all the tools developers are familiar with can be used—version control, diff, parameterization, reuse. The cost is a clear capability boundary: it can only do content that does not depend on live-action footage. It is not intended to replace diffusion models; it fills another piece of the puzzle: the images that should have been generated by programs but have kept being left for models to guess.

In video editing, re-roll should not be the default action. Code2Video at least offers another option.

Model: https://huggingface.co/sensenova/SenseNova-6.8-Code2Video

Code and skill: https://github.com/OpenSenseNova/SenseNova-6.8-Code2Video

Original post: https://x.com/mikenevermiss/status/2115727056373760003