SenseNova 6.8 Code2Video
SenseNova 6.8 Code2Video is an open-weight model from SenseTime for making short videos by writing and rendering code instead of generating pixels. It is a 27-billion-parameter fine-tune of Alibaba's Qwen3.8-27B with a 512K-token context window, released in late October 2026, with the Hugging Face weights under the Apache License 2.0, together with a production skill, a prompt-enhancement template and example production trajectories.[1][2] SenseTime describes it as "a lightweight open model for code-driven video creation" that "builds and renders videos by generating executable code," and the GitHub repository calls it "the first open-source model for code-driven video creation."[1][2] The model is meant to run inside an agent harness, and SenseTime's recommended harness is Claude Code.[1]
Because each element of a film is a named object in a program, a creator can ask for a change to timing, style or content and the model edits the same project and renders it again, leaving the parts that were not mentioned as they were. SenseTime presents this revision loop as the main difference from pixel-space video generation models.[2]
Background and naming
SenseTime publishes its SenseNova foundation models and related research releases under the sensenova organization on Hugging Face and the OpenSenseNova organization on GitHub, which describes itself as "SenseTime Foundation Model Sets."[15] Code2Video is the most recent model in that Hugging Face organization as of 29 October 2026.[1]
The "6.8" in the name matches the version number of SenseNova 6.8 Flash Lite, a lightweight multimodal agent model that SenseTime's website FAQ lists among its featured models as of 29 October 2026.[9] The Code2Video model card and README, however, describe the model as fine-tuned from Qwen3.8-27B and do not describe it as derived from SenseNova 6.8 Flash Lite.[1][2] The official name in the model card and README is written with spaces ("SenseNova 6.8 Code2Video"); the repository identifiers use hyphens (SenseNova-6.8-Code2Video).[1][2]
The model shares its name with Code2Video, an unrelated 2025 academic framework from the Show Lab at the National University of Singapore (see the related work section).[13]
Release
The Hugging Face repository and the GitHub repository were both created on 24 October 2026, and the model card was last modified on 27 October 2026.[1][2] The Chinese tech site 80aj.com reported the open-source release on 27 October, and the release was also discussed on the V2EX forum and, on 29 October, in English on X.[10][11][12]
| Released resource | Contents | Location |
|---|---|---|
| Model weights | BF16 safetensors in 12 shards, with config, tokenizer and processor files | Hugging Face[1] |
code-video skill | A workflow file (SKILL.md), eight craft guides and helper scripts for capture, audio and checking | GitHub[2][5] |
| Prompt-enhancement template | A template that turns a short request into a detailed production brief | GitHub[6] |
| Production trajectories | Message and tool-call logs of complete productions, with the rendered frames the model reviewed | Hugging Face[8] |
The Hugging Face model card states the license as Apache 2.0; as of 29 October 2026 the GitHub repository, which holds the skill and the prompt template, has no license file and its README does not state a license.[1][2] The GitHub README says the material is released "as a reference for researchers."[2]
Model
The released checkpoint is a vision-language model: its Hugging Face pipeline tag is image-text-to-text, and it keeps the vision encoder of its base model, which lets it look at rendered frames during production.[1][3] Its configuration uses the same Qwen3_5ForConditionalGeneration architecture class and qwen3_5 model type as the published Qwen3.8-27B repository.[1][4]
| Property | Value |
|---|---|
| Developer | SenseTime (SenseNova team)[1][2] |
| Base model | Qwen3.8-27B[1][2] |
| Parameters | 27,356,728,560 (BF16, per Hugging Face safetensors metadata)[1] |
| Layers | 64, in a repeating pattern of three linear-attention layers followed by one full-attention layer[3] |
| Hidden size | 5,120[3] |
| Vocabulary | 248,320 tokens[3] |
| Context length | 524,288 tokens (512K)[1][3] |
| Input | Text and images[1] |
| Output | Text (code, tool calls and reasoning)[1] |
| License | Apache 2.0[1] |
Qwen describes the layer layout of Qwen3.8-27B as 16 repeats of three Gated DeltaNet layers and one gated attention layer, with a native context of 262,144 tokens that can be extended with YaRN.[4] The only architectural setting that differs between the two config.json files is the rotary position embedding configuration: Code2Video switches it to YaRN scaling with a factor of 2.0 over 262,144 original positions, which gives the 524,288-token limit used in SenseTime's deployment commands.[3][4] Qwen's own model card recommends exactly that factor for applications whose typical context is 524,288 tokens.[4]
The base repository includes multi-token prediction (MTP) weights. The weight index of the Code2Video checkpoint lists no MTP tensors, which accounts for its slightly smaller parameter count than the 27,781,427,952 reported for Qwen3.8-27B.[4][18]
SenseTime says the model "supports a 512K context, handles context compaction well and can sustain sessions of over a thousand turns"; elsewhere in the same README it says the model sustains runs of "more than 10 hours and more than 2,000 turns."[2] These are the developer's own claims. The model card and README do not describe the fine-tuning data or method.
How it works
Why code
SenseTime's README argues that pixel-space generative video models sample a whole clip in one opaque process, which makes three things hard: physically correct motion, local edits (changing one detail means resampling the whole clip) and exact on-screen text, formulas and data.[2] In a code-rendered film every frame is a function of time, sizes and durations are explicit numbers, text is typeset verbatim, and "the same code always renders the same film."[2] The README also says code lets the model call other tools, such as image, video and 3D generation models, and describes the long-term goal as "a model that coordinates different generative models and creative tools through code."[1][2]
Production stages
SenseTime splits the production of one film into three stages:[1]
| Stage | Performed by | Output |
|---|---|---|
| 1. Prompt enhancement | A strong general model (SenseTime uses Claude Opus 5.5) with the enhance-prompt template | dense_prompt.md |
| 2. Production | SenseNova 6.8 Code2Video in Claude Code with the code-video skill | out/final.mp4, SPEC.md, VERIFY.md |
| 3. Revision (optional) | The creator writes notes; the model applies them in the same project | A new version of the film |
The README says the model can enhance prompts by itself "only at a basic level," and that running stage 1 with a strong model such as Claude improves the aesthetic quality of the result.[2] The enhancement template classifies requests by film type: explainers, art and visual effects, motion design and kinetic typography, product or brand promos, character or story animation, music-led films, and reproduction of a reference video. For a reference video, the template asks the producing model to analyze it frame by frame and rebuild it in code, and states that no pixels of the reference may be copied into the output.[6]
The code-video skill
The code-video skill is a set of agent skill instructions that SenseTime says is "fixed: the same version is used for every task."[2] It defines a code video as a program that draws a frame for a given time t, sampled at t = i / fps and encoded.[5] The skill tells the model to:
- write a
SPEC.mdwith format, exact on-screen copy, facts, assets and a list of must-have requirements before coding; - choose a rendering stack by content type;
- make every frame a pure function of
t, banning wall-clock timers, running animations and unseeded randomness; - capture frames and encode them with ffmpeg;
- review stills, contact sheets and short action strips as images, at least twice;
- verify the delivered file with a checking script and record pass or fail evidence for each requirement in
VERIFY.md.[5]
| Content | Default stack in the skill |
|---|---|
| 2D diagrams, math, particles, pixel art, painterly textures | Canvas 2D (in a browser or @napi-rs/canvas in Node) |
| Vector shapes, UI, kinetic type, logos | SVG or HTML/CSS on a paused timeline, or Remotion |
| LaTeX-heavy math | Manim (or KaTeX/MathJax rendered to SVG) |
| 3D scenes, products, camera moves | Three.js |
| Shaders, fields, post effects | WebGL / GLSL fragment shaders |
Browser-based films are captured by a script that opens the page in headless Chromium (through Playwright or Puppeteer), seeks each frame and pipes screenshots into ffmpeg, which encodes H.264 video.[5] The eight craft guides cover explainer diagrams, stylized looks, simulations and generative art, character and creature motion, Three.js scenes and cameras, motion graphics and promos, sound design and narration, and music and beat sync.[2][5] Sound is built from the same timeline as the picture and delivered as 48 kHz stereo AAC; the skill's dependencies include edge-tts as the default text-to-speech engine, faster-whisper to check narration, and librosa for beat tracking.[5][7]
Revision
After a film is delivered, the creator can give notes in the same session or resume it later with claude --continue; the model edits the existing project and renders a new version.[1][2] The README's example is a retro airport split-flap departures board that was restyled as a "bright pastel hotel lobby" in a second round and given a plane overhead and travellers with luggage in a third. SenseTime says each round took one short note and that "the copy, timing and flip mechanics we didn't ask to change stay exactly as they were."[2]
Deployment
SenseTime publishes Docker commands for serving the model with LightLLM, vLLM and SGLang. All three examples use four-way tensor parallelism and a 524,288-token context; the vLLM and SGLang examples use the qwen3 reasoning parser and the qwen3_coder tool-call parser.[1][2] The 12 BF16 weight shards total about 54.7 GB.[1]
For long runs SenseTime recommends these Claude Code settings:[1][2]
| Setting | Recommended value |
|---|---|
| Context window | 512K (CLAUDE_CODE_MAX_CONTEXT_TOKENS=524288) |
| Auto-compaction threshold | 80 percent |
| Maximum output | 64K tokens |
| Turns per run | 2,000 or more |
| Time per run | 10 hours or more |
The model card notes that Claude Code otherwise treats an unknown model as a 200K-context model, so the context size has to be set explicitly. Because the run executes shell commands without asking, SenseTime advises starting it in an isolated working environment.[1] A production also needs Node.js 18 or later, Python 3, ffmpeg and fonts for the video's language.[2]
Production trajectories
The file trajectories/all.openai.jsonl stores one complete production per line in OpenAI messages format, with the model's reasoning, tool calls and results, retries and compaction summaries. Images appear as placeholders, and the model card says all 357 figures are included in a separate folder. Private paths and hosts are anonymized, so the recorded commands cannot be replayed directly.[1] As of 29 October 2026 the file contains six productions, including films titled "Fourier Light Painting" and "Signal / Noise / Shape" that also appear in the showcase.[8]
Showcase
SenseTime says its 90-second launch film, and every video shown in it, was made by the model.[2] The README lists nine showcase films: "Can a Forest Hide a Landslide?", "Fourier Light Painting", "Signal / Noise / Shape", "Paper & Ink", "Clock Clock", "Golden Spiral and Sunflower", "Paper Sea", "A to Z" and "Robot and the Night Lamp." The repository's showcase folder also holds three further films, two of them product promos.[2] Chinese coverage also described an official example in which the model rebuilt a product demo video from the MP4 file alone, then renamed the product to KITE and changed the canvas to light green.[10][11]
Limitations
The README lists these limitations:[2]
- Scope. The model only makes videos that can be written as code, such as motion graphics, explainers, stylized 2D and 3D scenes and procedural audio. It does not generate photorealistic footage.
- Cost. A production run takes hours and a long context, and results depend on the harness configuration.
- Evaluation. SenseTime reports only qualitative examples. It says it has not built a systematic benchmark for code-driven video and considers one an open problem.
- Gap to frontier models. SenseTime says the model "is clearly behind the strongest proprietary models on visual polish and long-horizon reliability."
- Prompt enhancement. The model's own prompt enhancement is basic.
Reception
Within days of the release, community members published GGUF quantizations of the model on Hugging Face, including an MTP-GGUF build and a GSQ-RCO GGUF build.[16][17] Chinese reports said SenseTime planned INT4 and INT8 quantized versions.[10][11] A V2EX post compared its output to videos made with Claude Opus 5.5 and noted that "reproducing" a reference video means writing a new code implementation, not recovering the original project files.[11]
Related work
The approach belongs to a line of work that treats video as a program rather than as pixels. The most direct namesake is Code2Video: A Code-centric Paradigm for Educational Video Generation by Yanzhe Chen, Kevin Qinghong Lin and Mike Zheng Shou of Show Lab, first posted to arXiv on 1 October 2025.[13] It is an agent framework, not a trained model: a Planner structures lecture content, a Coder turns it into executable Manim code with scope-guided auto-fix, and a Critic uses a vision-language model with "visual anchor prompts" to refine layout. Its authors also built the MMMC benchmark and the TeachQuiz metric, and report a 40 percent improvement over direct code generation.[13] The paper appeared in the proceedings of the 43rd International Conference on Machine Learning (ICML 2026).[14]
The two projects share the code-rendered approach, but SenseNova 6.8 Code2Video is a separately trained model from a different team, and its README does not cite the Show Lab paper.[2][13] Both contrast with pixel-space text-to-video generators such as Sora and Veo, which synthesize the frames of a clip directly rather than writing a program that renders them.
Continue reading
- SenseTimeSenseTime (Chinese: 商汤科技; pinyin: Shāngtāng Kējì) is a partly state-owned, publicly traded artificial intelligence company headquartered in Hong Kong that grew into China's largest AI software company by…Mentioned 22 times above
- Qwen3.8Qwen3.8 is the name Qwen uses for a model generation that includes hosted services and downloadable checkpoints.Mentioned 6 times above
- Claude CodeClaude Code is an agentic software development product made by Anthropic. It uses models from the Claude family to inspect software projects, propose and apply changes, run development tools, and report…Mentioned 5 times above
- Fine-tuningFine-tuning is the process of adapting a pretrained machine-learning model by continuing to optimize some or all of its parameters on data selected for a target task, domain, behavior, or population.Mentioned twice above
- Context windowA context window is the finite token sequence that a language model can process for one invocation.Mentioned twice above
- SenseNovaSenseNova (Chinese: 日日新) is the family of foundation models developed by the Hong Kong-listed AI company SenseTime.Mentioned 10 times above
More to read
Topics:AI Code GenerationChinese AIOpen Source AIVideo Generation
References
- ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11 ^12 ^13 ^14 ^15 ^16 ^17 ^18 ^19 ^20 ^21 ^22 ^23 ^24 ^25 ^26SenseTime (sensenova). "SenseNova 6.8 Code2Video" model card. Hugging Face, accessed 29 October 2026. huggingface.co/...SenseNova-6.8-Code2Video
- ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11 ^12 ^13 ^14 ^15 ^16 ^17 ^18 ^19 ^20 ^21 ^22 ^23 ^24 ^25 ^26 ^27OpenSenseNova. "SenseNova-6.8-Code2Video" README. GitHub, accessed 29 October 2026. github.com/...SenseNova-6.8-Code2Video
- ^1 ^2 ^3 ^4 ^5 ^6SenseTime (sensenova). "config.json" for SenseNova-6.8-Code2Video. Hugging Face. huggingface.co/...config.json
- ^1 ^2 ^3 ^4 ^5Qwen. "Qwen3.8-27B" model card and config. Hugging Face. huggingface.co/...Qwen3.8-27B
- ^1 ^2 ^3 ^4 ^5 ^6OpenSenseNova. "code-video" skill (SKILL.md). GitHub. github.com/...SKILL.md
- ^1 ^2OpenSenseNova. "Enhance prompt" template. GitHub. github.com/...enhance_prompt.md
- ^OpenSenseNova. "requirements.txt" for the code-video skill. GitHub. github.com/...requirements.txt
- ^1 ^2SenseTime (sensenova). "trajectories/all.openai.jsonl." Hugging Face. huggingface.co/...all.openai.jsonl
- ^SenseTime. "Frequently Asked Questions." SenseTime, accessed 29 October 2026. sensetime.com/...faq
- ^1 ^2 ^3"商汤开源27B代码生视频模型Code2Video：先写代码再渲染，自然语言可迭代修改." 80aj.com, 27 October 2026. 80aj.com/...code2video-27b-model
- ^1 ^2 ^3 ^4"商汤发布 SenseNova 6.8 Code2Video：把动态字体、图形动画和知识讲解写成代码." V2EX, October 2026. v2ex.com/...1252373
- ^@Chinazhidx. Post on X about SenseNova 6.8 Code2Video, 29 October 2026. x.com/...2115664926245064936
- ^1 ^2 ^3 ^4Chen, Yanzhe; Lin, Kevin Qinghong; Shou, Mike Zheng. "Code2Video: A Code-centric Paradigm for Educational Video Generation." arXiv:2510.01174, 1 October 2025. arxiv.org/...2510.01174
- ^Chen, Yanzhe; Lin, Kevin Qinghong; Shou, Mike Zheng. "Code2Video: A Code-centric Paradigm for Educational Video Creation." Proceedings of the 43rd International Conference on Machine Learning, PMLR 306:15370-15396, 2026. proceedings.mlr.press/...chen26bx
- ^OpenSenseNova. GitHub organization page. github.com/OpenSenseNova
- ^Abiray. "SenseNova-6.8-Code2Video-MTP-GGUF." Hugging Face, 27 October 2026. huggingface.co/...SenseNova-6.8-Code2Video-MTP-GGUF
- ^OliviaRossi. "SenseNova-6.8-Code2Video-GSQ-RCO-GGUF." Hugging Face, 28 October 2026. huggingface.co/...Nova-6.8-Code2Video-GSQ-RCO-GGUF
- ^SenseTime (sensenova). "model.safetensors.index.json" for SenseNova-6.8-Code2Video. Hugging Face. huggingface.co/...model.safetensors.index.json
Improve this article
Add missing citations, update stale details, or suggest a clearer explanation. Every suggestion is reviewed for sourcing before it goes live.
1 revision · v1 · 2,519 words · full history
Fact-checks are independent of edits: a reviewer re-verifies the article against its sources and stamps the date. How we verify
Research and drafting on this wiki are AI-assisted, under named human editorial standards. How AI is used here
Reviewer note: Model specifications, license, release contents and showcase material checked against the Hugging Face model card and config, the GitHub repository and Chinese press coverage, October 2026.
Cite this page: AI Wiki. "SenseNova 6.8 Code2Video." aiwiki.ai, updated 29 Oct 2026, fact-checked 29 Oct 2026. CC BY 4.0. https://aiwiki.ai/wiki/sensenova_6_8_code2video