SenseNova 6.8 Code2Video

SenseNova 6.8 Code2Video is an open-source code-driven video creation model from SenseTime, with the tagline "Every frame is code." Instead of painting pixels directly, it writes executable code and renders it into video, then reviews the frames and revises the code, so every frame has readable, editable source. SenseTime's official account @SenseTime_AI introduced it on 2026-10-28.

Model: huggingface.co/sensenova/SenseNova-6.8-Code2Video · Skill and prompts: GitHub OpenSenseNova/SenseNova-6.8-Code2Video · License Apache-2.0 · 27B parameters · 512K context.

Key points

1. Base: fine-tuned from Qwen3.8-27B for long-horizon, agent-style film production; SenseTime says it sustains runs of 10+ hours and 2,000+ turns.
2. What it makes: currently short videos built entirely in code, such as math explainers, paper-craft films, isometric worlds, kinetic type, data stories, character animation and product promos. SenseTime reports good visual composition and motion choreography in motion design in preliminary results.
3. Released resources: weights, config and tokenizer; example production trajectories with reasoning and tool calls (OpenAI messages format, 357 figures); and the code-video production skill on GitHub (workflow, craft guides, rendering tools).
4. Deployment: the model card has Docker recipes for LightLLM, vLLM and SGLang (examples use 4-way tensor parallelism and a 512K context).
5. Long-term goal: use code as a common language that chains image, video, 3D and editing models and tools, and extend the model to editing, compositing, animation and 3D content.

Why it's tracked

A new, clearly named open-source model formally launched by a major vendor in this scan window, with public weights and an official announcement, opening a distinct "code-to-video" niche.

Production flow (three official stages)

1. Prompt enhancement: use a strong general model (SenseTime uses Claude Opus 5.5) with the repo's Enhance prompt template to expand your idea into dense_prompt.md.
2. Production: serve the model with vLLM / SGLang / LightLLM, point Claude Code at your endpoint, copy the code-video skill into .claude/skills/code-video, and run claude -p ... --max-turns 2000. Outputs are out/final.mp4, SPEC.md and VERIFY.md.
3. Revision (optional): write your notes and run claude -p --continue in the same folder to get a new version.

Tips

• Claude Code treats unknown models as 200K context, so set CLAUDE_CODE_MAX_CONTEXT_TOKENS=524288; auto-compaction at 0.8 and 64K max output are recommended.
• Runs execute shell commands without asking, so use an isolated environment.
• A 27B model at 512K context needs multiple GPUs (the examples use 4).