Apache-2.0multimodalRuns locally27B (card; safetensors 27.36B)
SenseNova 6.8 Code2Video: Specs, License & How to Run
SenseTime Apache-2.0 27B code-driven video (Qwen3.8-27B fine-tune). Not a pixel DiT.
- License
- Apache-2.0
- Parameters
- 27B (card; safetensors 27.36B)
- HF likes
- 45 (2026-10-29)
- First seen
- 2026-10-24
TL;DR. SenseNova 6.8 Code2Video is a multimodal model from SenseTime (SenseNova). Parameters: 27B (card; safetensors 27.36B).License: Apache-2.0. It is listed under Runway, Claude Code.
| Developer | SenseTime (SenseNova) |
|---|---|
| Type | multimodal |
| Architecture | Card and GitHub: fine-tuned from Qwen3.8-27B for code-driven video. HF model_type qwen3_5; architecture Qwen3_5ForConditionalGeneration. pipeline_tag image-text-to-text. text_config: 64 layers, hidden 5120, vocab 248320, max_position_embeddings 262144. The model writes executable code that renders short videos; it is not a pixel-direct video generator like LTX-2.5. |
| Parameters | 27B (card; safetensors 27.36B) |
| Modalities | text, image, code, video |
| Languages | en, zh |
| Context | 512K serving target (README); 262,144 in config.json text_config |
| License | Apache-2.0 — Hugging Face cardData.license is apache-2.0 and the HF README states “Released under the Apache License 2.0” (opened 2026-10-29). There is no LICENSE sibling on the weight repo (resolve/main/LICENSE 404) and no LICENSE file in the GitHub repo root; GitHub’s license API field is empty. Not legal advice. |
| Hardware | Not yet verified — HF usedStorage is 54,903,399,420 bytes (~51.1 GiB). safetensors.parameters sum to 27,356,728,560. Card deploy examples use tensor-parallel-size 4 (LightLLM / vLLM / SGLang) with --max-model-len 524288. Peak VRAM is not independently measured. Do not invent a single-GPU footprint. (not independently tested) |
| Hugging Face likes | 45 (as of 2026-10-29) |
| Model card | https://huggingface.co/sensenova/SenseNova-6.8-Code2Video |
| First seen | 2026-10-24 |
SenseNova 6.8 Code2Video is SenseTime’s ungated 27B checkpoint for code-driven short video. Official hosts we opened on 2026-10-29: the card sensenova/SenseNova-6.8-Code2Video (created 2026-10-24T09:06:28Z; last modified 2026-10-27T12:27:33Z; 45 likes / 66 downloads) and github.com/OpenSenseNova/SenseNova-6.8-Code2Video (79 stars / 10 forks; created 2026-10-24T07:58:58Z; last push 2026-10-27T14:15:35Z). API gated was false. Hugging Face pipeline_tag is image-text-to-text. Card tags include code-video, video-generation, agent, claude-code, qwen3.8.
This page hangs off Runway alternatives (video-creation search) and Claude Code alternatives (the README’s recommended harness). Those mappings are editorial. This is not a pixel-direct video model and not a Claude Code clone. We have no sourced quality comparison against Runway, Sora, or Claude Code.
Code-driven vs pixel generative
| Job | On this site |
|---|---|
| Pixel-direct audio-video DiT | LTX-2.5 on Runway alternatives |
| HTML/CSS → deterministic MP4 | HyperFrames on HeyGen alternatives |
| LLM writes code, renders frames, reviews stills | This card |
Copied from the GitHub Limitations section: the model makes videos that can be written as code — motion graphics, explainers, stylized 2D and 3D scenes, procedural audio — and does not generate photorealistic footage. Evaluation is qualitative examples only; the README says there is no systematic benchmark yet. Those sentences are upstream.
The Hugging Face card and GitHub README both state the weights are fine-tuned from Qwen3.8-27B. Same-site base card: Qwen3.8-27B.
What we verified (2026-10-29)
| Field | Value | Source |
|---|---|---|
| Host | huggingface.co/sensenova/SenseNova-6.8-Code2Video (not gated) | HF API |
| Created / last modified | 2026-10-24T09:06:28Z / 2026-10-27T12:27:33Z | same |
| Likes / downloads | 45 / 66 | same |
| License tag | apache-2.0 | HF cardData |
| LICENSE on the weight repo | 404 | HF siblings |
| GitHub LICENSE file / API | none / empty | GitHub contents + license API |
| Pipeline tag | image-text-to-text | HF API |
| Architecture | qwen3_5 / Qwen3_5ForConditionalGeneration | config.json |
| Params (card) | 27B | card + GitHub |
| safetensors.parameters | 27,356,728,560 | HF API |
| usedStorage | 54,903,399,420 bytes (~51.1 GiB) | same |
| text_config | 64 layers / hidden 5120 / vocab 248,320 | config.json |
| config.json context | 262,144 | text_config.max_position_embeddings |
| Serving context named | 512K (--max-model-len 524288) |
GitHub / HF deploy examples |
| GitHub stars / forks | 79 / 10 | GitHub API |
| Base (card + README) | Qwen3.8-27B fine-tune | HF README + GitHub |
runsLocally: true means the weights are downloadable. The documented serve path uses tensor parallel 4. That is a vendor recipe, not a forkalt VRAM measurement.
How to run (from official README, not yet tested)
Card/GitHub name LightLLM, vLLM, and SGLang, each with --tp 4 / --tensor-parallel-size 4 and max length 524288. vLLM example flags copied from the card: --reasoning-parser qwen3, --enable-auto-tool-choice, --tool-call-parser qwen3_coder.
The recommended harness is Claude Code pointed at that endpoint (context 512K, compaction 0.8, max output 64K, long --max-turns). Production uses the code-video skill in the GitHub repo. Prompt enhancement on the README is a separate strong general model (they name Claude Opus 5.5). forkalt has not run those commands.
License
Hugging Face license: apache-2.0. HF README: “Released under the Apache License 2.0.” We did not find a LICENSE blob on the weight repo or in the GitHub repo root on 2026-10-29. GitHub’s license API field is empty. Treat the grant as the card tag plus README sentence until a file appears. Not legal advice.
What we did not invent
- No vendor numeric video benchmark is copied here; the README says evaluation is qualitative only.
- Peak VRAM, wall-clock for a full film, and “10 hours / 2,000 turns” harness claims stay upstream.
- We do not rank this card against LTX-2.5, HyperFrames, Runway, or Claude Code.
Related: Pika H3 Flex (pixel/audio-video adapters; local 480p/768p) — Pika H3 Flex vs SenseNova 6.8 Code2Video — LTX-2.5 (pixel DiT) — HyperFrames (HTML→video) — Qwen3.8-27B (base card) — vLLM — Claude Code alternatives — Runway alternatives. Catalog: home. Editorial: About.
Deploy options
- local: Hugging Face weights (BF16)
- local: GitHub skill / deploy docs
Comparisons
FAQ
Does this generate photorealistic video like LTX-2.5 or Runway?
No. The GitHub limitations section says the model makes videos that can be written as code (motion graphics, explainers, stylized 2D/3D, procedural audio) and does not generate photorealistic footage. LTX-2.5 is a pixel-direct audio-video DiT on Runway alternatives. HyperFrames is HTML→MP4. Different jobs.
Is it really a Qwen3.8-27B fine-tune?
Yes, as stated on the Hugging Face card (“fine-tuned from Qwen3.8-27B”) and the GitHub README (“built on Qwen3.8-27B”). config.json model_type is qwen3_5, matching this site’s Qwen3.8-27B identification page. Safetensors total 27,356,728,560 versus 27,781,427,952 on Qwen/Qwen3.8-27B.
Are the weights Apache-2.0?
cardData.license is apache-2.0 and the HF README says Apache License 2.0. We did not find a LICENSE file on the Hugging Face card or the GitHub repo root (2026-10-29). GitHub’s license API field is empty. Not legal advice.
Does it replace Claude Code or Runway?
No. The recommended harness is Claude Code pointed at a self-hosted endpoint. Runway is a commercial pixel video studio. This page identifies an open 27B code-video checkpoint plus a skill. We have no sourced bake-off.
Sources & verification
- https://huggingface.co/sensenova/SenseNova-6.8-Code2Video
- https://github.com/OpenSenseNova/SenseNova-6.8-Code2Video
- https://huggingface.co/Qwen/Qwen3.8-27B
- https://huggingface.co/Lightricks/LTX-2.5
License notes are not legal advice. Read the original LICENSE. Trademark & disclaimer.