SenseTime's "code-to-video" model fine-tuned from Qwen3.8-27B: Used within Claude Code alongside the official code-video skill, it writes programs to render frames one by one, reviews rendering screenshots, iterates repeatedly, and finally delivers final.mp4; it creates dynamic graphics, mathematical explanations, stylized 2D / 3D scenes, and procedural audio effects expressible via code, rather than generating photorealistic visuals.
Upvotes 61 · Likes 61 · Downloads 266
Reasons
- Unlike diffusion video models, every element in the visuals is a named object in the code: text, formulas, and chart data are character-perfect; changing size, timing, or motion trajectories requires modifying only a few lines of code while the rest remains unchanged; after delivery, revisions can be implemented by re-rendering within the same project.
- Officially claimed to run continuously for over 10 hours and 2,000 rounds, with 512K context; the complete production trajectory (conversations, tool calls, and 357 review screenshots) has been released, facilitating research on long-horizon agents.
- Released on 24 10, under the Apache-2.0 license, and not included in previous issues; the card includes a 90-second release short film, which the official team states was produced by this model.
Usage
- Deployment: LightLLM, vLLM, or SGLang with four-card tensor parallelism (--tp 4 / --tensor-parallel-size 4), context set to 524288; for vLLM / SGLang, use the qwen3 reasoning parser and the qwen3_coder tool parser.
- Point Claude Code to your own endpoint: In private settings, set ANTHROPIC_BASE_URL, CLAUDE_CODE_MAX_CONTEXT_TOKENS=524288, CLAUDE_AUTOCOMPACT_PCT_OVERRIDE=80, and CLAUDE_CODE_MAX_OUTPUT_TOKENS, with --max-turns 2000; copy the repository's skills/code-video directory as-is to .claude/skills/code-video in the task directory. Additionally, Node.js 18+, Python 3, ffmpeg, and Playwright Chromium are required.
- Three steps: First, use a strong model to expand a single requirement into dense_prompt.md using the official enhance_prompt template; then, have this model produce the video according to the skill (out/final.mp4, SPEC.md, VERIFY.md); finally, use --continue to write revision notes for iteration.
Risks
- Only qualitative examples are provided; the official team states that systematic evaluations have not yet been conducted and acknowledges that visual completeness and long-horizon reliability lag significantly behind the strongest closed-source models.
- Producing a single video requires hours of agent time and a very long context; the 27B model is deployed on four cards; official examples use acceptEdits mode, allowing the agent to execute commands without confirmation, so it must be run in an isolated environment.
- It cannot generate photorealistic videos; the official team recommends delegating the prompt expansion step to a stronger general-purpose model.