SenseNova-Loop-9B-Preview vs SenseNova 6.8 Code2Video

Official SenseNova-Loop-9B-Preview looped-decoder and two-stream mHC overview from the Hugging Face card — sensenova/SenseNova-Loop-9B-Preview assets/loop-mhc-overview.jpg. Source · License: Apache-2.0 (card tag / GitHub LICENSE).

Official SenseNova 6.8 Code2Video title banner from the Hugging Face card — sensenova/SenseNova-6.8-Code2Video media/readme/title.gif (first frame). Source · License: Apache-2.0 (card tag).

Verdict. Pick SenseNova-Loop-9B-Preview if you want SenseTime’s Apache-2.0 ~9B looped multimodal chat/agent checkpoint (text+image in, text out, 256K training context, custom Transformers code). Pick SenseNova 6.8 Code2Video if you want the 27B Qwen3.8-27B fine-tune that writes code to render short motion graphics / explainers — not photorealistic footage. Same SenseNova family. We have no sourced bake-off. This pairing does not replace either entity page.

Compare SenseNova-Loop-9B-Preview and SenseNova 6.8 Code2Video.

Dimension	SenseNova-Loop-9B-Preview	SenseNova 6.8 Code2Video	Source
License	Apache-2.0 (card tag + GitHub LICENSE file)	Apache-2.0 (card tag + README; no LICENSE blob on 2026-10-29)	Hugging Face / GitHub (2026-11-03 / 2026-10-29)
Gating	No	No	Hugging Face API
Developer	SenseTime (SenseNova)	SenseTime (SenseNova)	Official cards
Job	Multimodal chat / agent (text+image → text)	LLM writes code that renders short video	Official cards
Parameters	~9B total; ~14.6B inference params (card); HF safetensors 9.66B	27B (card); safetensors 27,356,728,560	Official cards / HF API
Architecture	sensenova_loop; layers 4–27 applied twice; two-stream mHC	qwen3_5; fine-tuned from Qwen3.8-27B	Official cards / config.json
Context length	262,144 (card training + config)	512K serving target (README); 262,144 in config.json	Official cards
Serving named	Transformers + trust_remote_code; LightLLM tp 4; vLLM/SGLang via named PRs	LightLLM / vLLM / SGLang with tensor-parallel 4	Official cards
Hugging Face likes	184 (2026-11-03)	45 (2026-10-29)	Hugging Face API
Quality vs ChatGPT / Runway	Not independently tested	Not independently tested	n/a

When to choose which

Looped multimodal chat / long-context agent card: SenseNova-Loop-9B-Preview — Text+image in, text out. Not a video generator.
Code-driven motion graphics / explainers: SenseNova 6.8 Code2Video — GitHub limitations section. Not photorealistic DiT.
Qwen3.8-27B fine-tune identification: SenseNova 6.8 Code2Video — Loop-9B copies qwen3_5_text as a config label only. It does not claim a Qwen3.8-27B fine-tune on the README we opened.
Pixel-direct Runway-class video already on this site: neither — Use LTX-2.5 or Pika H3 Flex on Runway alternatives.

Read the entity pages: SenseNova-Loop-9B-Preview and SenseNova 6.8 Code2Video. Loop hangs off ChatGPT alternatives and closed LLM APIs. Code2Video hangs off Runway alternatives and Claude Code alternatives.

Why this comparison exists

Both cards use the SenseNova name. They are not the same job. Loop-9B-Preview is a ~9B looped multimodal chat/agent checkpoint (released 2026-10-30 per the card news line). Code2Video is a 27B code-driven video fine-tune of Qwen3.8-27B. This page exists so neither URL swallows the other.

Sourced snapshot (Loop 2026-11-03 / Code2Video 2026-10-29)

Field	SenseNova-Loop-9B-Preview	SenseNova 6.8 Code2Video
Host	huggingface.co/sensenova/SenseNova-Loop-9B-Preview (not gated)	huggingface.co/sensenova/SenseNova-6.8-Code2Video (not gated)
HF likes	184 (2026-11-03)	45 (2026-10-29)
License	apache-2.0 card tag; GitHub LICENSE file	apache-2.0 card tag; no LICENSE blob (2026-10-29)
Size	~9B / ~14.6B inference (card)	27B (card)
Job	text+image → text	code-driven short video
Architecture	looped decoder + two-stream mHC	Qwen3.8-27B fine-tune
Context	262,144	512K serving target; 262,144 in config
Quality vs ChatGPT / Runway	Not independently tested	Not independently tested

How to choose

Need looped multimodal chat (~9B, Apache-2.0 card tag): SenseNova-Loop-9B-Preview.
Need code-driven explainers / motion graphics: SenseNova 6.8 Code2Video.
Need a pixel-direct audio-video DiT: LTX-2.5.
Need local 480p/768p pixel/audio-video adapters: Pika H3 Flex — Pika H3 Flex vs SenseNova 6.8 Code2Video.
Need Qwen’s 27B native VLM: Qwen3.8-27B.

FAQ

Which one generates video?

Code2Video writes code that renders short motion graphics. Loop-9B-Preview is a multimodal language model that emits text. Pixel-direct video on this site is LTX-2.5 or Pika H3 Flex, not this pair.

Are both Apache-2.0?

Both cards tag apache-2.0. Loop-9B has an Apache-2.0 LICENSE on OpenSenseNova/SenseNova-Loop. Code2Video had no LICENSE blob on 2026-10-29. Not legal advice.

Why this pairing?

Same SenseNova family, two different jobs. Searchers who land on one name should not replace the other page. Identification, not a quality rank.

Can stock vLLM serve Loop-9B?

The Loop card names a specific vLLM pull request and trust_remote_code for Transformers. Code2Video’s GitHub names LightLLM / vLLM / SGLang with tp 4. forkalt has not served either.

Sources & verification

Last verified 2026-11-03 · Last updated 2026-11-03· Edited by forkalt editors · Editorial policy

License notes are not legal advice. Read the original LICENSE. Trademark & disclaimer.