Introduction

What if the thing you got back from a video model wasn't just an MP4, but the source code for it? That's the pitch behind SenseNova 6.8 Code2Video, which SenseTime released as open weights on Hugging Face and formally launched on October 28. It's a 27B model fine-tuned from Qwen3.8-27B, licensed Apache 2.0, with a 512K context window.

Instead of sampling frames like a diffusion model, Code2Video writes executable code that draws every frame as a function of time, renders it, looks at the result, and revises. You run it today on your own GPUs through vLLM, SGLang, or LightLLM, driven by an agent harness (SenseTime recommends Claude Code) and a free code-video skill from the GitHub repo.

What shipped

  • Weights: sensenova/SenseNova-6.8-Code2Video on Hugging Face, plus community GGUF quants already appearing.
  • Skill: code-video on GitHub, with eight craft guides (explainers, stylized looks, simulations, characters, 3D scenes, motion graphics, sound and narration, music sync) and rendering and checking tools.
  • Prompt template: an "enhance prompt" brief that turns a one-line idea into a dense production spec.
  • Trajectories: full production logs for the showcase films, including tool calls and 357 review images.
  • Launch film: a 90-second piece with sound, made entirely by the model.

How it works

A film comes together in three stages:

  • Enhance: a strong general model expands your idea into dense_prompt.md using SenseTime's template.
  • Produce: Code2Video runs inside the agent harness with the code-video skill. It writes a spec, codes the scenes, renders contact sheets, reviews them as images, and delivers out/final.mp4 with SPEC.md and VERIFY.md.
  • Revise: you leave notes ("make the title bigger", "slow the chart") and it edits the same project and re-renders only what changed.

SenseTime says the model can sustain runs of more than 10 hours and 2,000+ turns. That's the honest cost: this is a patient agent, not a ten-second clip button.

What it's good at (and not)

Think motion design, not cinema. The showcase covers math explainers, paper-craft films, isometric worlds, kinetic type, data stories, light character animation, and product promos. Because text is typeset rather than generated, titles, formulas, and chart values come out exactly right, which is the classic weak spot of pixel video models.

SenseTime is upfront about the limits. It doesn't generate photorealistic footage, it trails the strongest closed models on polish and long-run reliability, and there's no formal benchmark yet, only qualitative examples.

Try it

The reference deploy splits the model across four GPUs. Here's the vLLM server from the model card:

docker run --rm --gpus all --ipc=host -p 8000:8000 \
  -e VLLM_ALLOW_LONG_MAX_MODEL_LEN=1 \
  -v /path/to/SenseNova-6.8-Code2Video:/model \
  vllm/vllm-openai:latest \
    --model /model \
    --tensor-parallel-size 4 \
    --max-model-len 524288 \
    --reasoning-parser qwen3 \
    --enable-auto-tool-choice \
    --tool-call-parser qwen3_coder \
    --host 0.0.0.0 --port 8000

Then drop the skill into a fresh task folder:

TASK=/absolute/path/to/a-new-video-task
git clone https://github.com/OpenSenseNova/SenseNova-6.8-Code2Video.git
mkdir -p "$TASK/.claude/skills"
cp -r SenseNova-6.8-Code2Video/skills/code-video "$TASK/.claude/skills/"

You'll also need Node.js 18+, Python 3, ffmpeg, Playwright's Chromium, and fonts for your video's language. Point your harness at the endpoint with the context window set to 512K, since unknown models otherwise get treated as 200K. Run it in an isolated folder, because the agent executes shell commands on its own.

Why it matters

Pixel video models are getting gorgeous, but editing them still means re-rolling and praying the parts you liked survive. Code2Video flips that: every element is a named object, so a note becomes a precise code edit. SenseTime also pitches code renders as fast previs that a diffusion model could later paint over, and that's the hybrid we'd love to see in ComfyUI next.

Original Source

Conclusion

Code2Video isn't trying to out-render Seedance or H3. It's trying to make video you can open, read, and edit like a document. If you make explainers, lyric videos, or data stories, this is a fascinating open toy, as long as you have the GPUs and the patience for a long agent run.

—Aurelia ♡