Overview
SenseNova 6.8 Code2Video is SenseTime's open-source video model, released on October 28, 2026. The approach differs from frame-by-frame diffusion: the model writes code, and the code renders the picture. Every frame is the output of an executable program, which makes the visuals precisely controllable and editable in place.
The model has 27 billion parameters, is fine-tuned from Qwen3.8-27B, carries a 512K context, runs in BF16, and ships under Apache 2.0 with weights on Hugging Face and ModelScope. Its loop is write code, render, look at frames, revise. SenseTime reports these runs can continue for more than ten hours and over 2,000 turns.
The published pipeline has three stages: enhance the prompt with a strong general model into dense_prompt.md, produce out/final.mp4, SPEC.md and VERIFY.md inside Claude Code with the code-video skill, then keep iterating inside the same project on the creator's notes.
Key Features
- Code is the picture: The output is runnable code rather than a tensor of pixels. Changing a parameter changes one part of the frame, with no variation from regenerating the whole clip.
- Ten hours, 2,000 turns: The recommended environment is configured for long-horizon work: 512K context, auto-compaction at 80 percent, 64K output cap, 2,000 turns and a ten-hour timeout. The model judges its own frames and revises.
- Verifiable artefacts: Each production writes final.mp4, SPEC.md and VERIFY.md. The spec records what the piece should be, the verification records self-checks, so revisions can be argued against a document.
- Bundled code-video skill: The skill directory carries workflow rules, craft guides and rendering tools, copied into .claude/skills/code-video. The craft guides cover composition and motion design, not just a prompt.
- Full trajectories published: The trajectories directory holds each demo's complete generation trace with 357 process images, alongside the prompt enhancement template in the same repository, so the chain can be reproduced.
- Three serving paths: LightLLM, vLLM and SGLang all have official launch arguments at tensor parallel four with a 524288 context. Transformers and SDK call examples are included as well.
Use Cases
- Math explainers and data stories, where chart and formula elements are far more controllable as code than as pixels.
- Kinetic typography and title sequences, with motion curves written directly and pacing changed without regenerating.
- Stylised shorts such as paper craft or isometric worlds, where composition and transitions are parameterised.
- Bulk versions of product promos, changing copy and colour while the structure holds.
- Creative projects that must ship source code, since the output is readable code by design.
- Workflows chaining generation, editing and 3D, because the model can call external tools from code.
Pros
- Apache 2.0 with weights on both Hugging Face and ModelScope.
- 27 billion parameters and 512K context, runnable at tensor parallel four.
- Visuals determined by code, so one change affects one place.
- Runs for over ten hours and 2,000 turns, suited to repeated revision.
- Writes SPEC.md and VERIFY.md so revisions have a reference.
- Trajectories and prompt templates are public, making the chain reproducible.
Pricing
Open weights under Apache 2.0, free to download and use commercially, with self-hosting costing only hardware. SenseTime recommends tensor parallel four; vLLM needs VLLM_ALLOW_LONG_MAX_MODEL_LEN=1 and SGLang needs SGLANG_ALLOW_OVERWRITE_LONGER_CONTEXT_LEN=1. The production pipeline is designed around Claude Code, whose model costs are billed separately.
Summary
Generating video as code buys controllability. Diffusion resamples and drifts between runs; rendered code does not, so a revision is a line change in the same project. The cost is a host that executes shell commands, and SenseTime notes the runtime runs them without confirmation, so isolate the workspace. The release gives qualitative results only and no quantitative benchmark, so compare against the published trajectories before committing.
Version History
- SenseNova 6.8 Code2Video open-sourced: video rendered from code (2026-10-28): SenseTime released and open-sourced SenseNova 6.8 Code2Video, 27 billion parameters, 512K context, BF16, Apache 2.0, fine-tuned from Qwen3.8-27B. It generates executable code and renders the frames, inspecting and revising them, with runs reported past ten hours and 2,000 turns. The three-stage pipeline covers prompt enhancement, production inside Claude Code with the code-video skill, and iteration in the same project, emitting final.mp4, SPEC.md and VERIFY.md. Weights are on Hugging Face and ModelScope, and trajectories include 357 process images.