AIToday
Large Language ModelsOpen-Source AIAI WeeklyPublished: Nov 18, 2026, 04:00 JST

SenseNova-Loop-9B drops on Hugging Face under Apache 2.0

Send on LINE
SenseNova-Loop-9B drops on Hugging Face under Apache 2.0

3 Key Points

  1. What happened

    SenseNova published SenseNova-Loop-9B on Hugging Face under Apache 2.0, a multimodal decoder that loops 24 middle layers 56 times per forward pass, about 14.6B inference-equivalent parameters, with 512K context.

  2. Why it matters

    A small model that reuses its middle layers can carry far more effective depth than its size suggests, while the Apache 2.0 license lets companies deploy it without permission.

WHO IT HITSTeams that run their own models without paying for a license can now pick up a 512K-context multimodal model directly from Hugging Face. Serving-engine maintainers face PRs to add vLLM and SGLang support.

Not sure about something? Ask the AI

Questions and answers are published on this page.

Summaries like this, every morning.

Context & Analysis

SenseNova-Loop-9B sits at an unusual point on the size spectrum. Its physical layer count is 32 and its nominal size is 9B, yet because it reuses 24 of those middle layers 56 times per forward pass, its inference-equivalent parameter count comes to about 14.6B. The design targets effective depth by looping rather than by stacking more layers.

The release terms matter as much as the architecture. Publishing under Apache 2.0 on Hugging Face means the weights carry no permission step for commercial deployment — a different posture from models shared under narrower licenses. The 512K context and image input put it in the multimodal category rather than plain text.

Practical adoption looks tied to serving support: vLLM and SGLang compatibility is described as landing via open PRs, so the tooling around the model is still in motion. SenseNova also reports benchmark comparisons against Qwen3.8-27B, which suggests it is positioning a comparatively compact model against a larger one.

FAQ
What does the 'loop' in Loop-9B actually mean?
The model reuses 24 middle layers to run its 32 physical layers 56 times per forward pass, reaching roughly 14.6B inference-equivalent parameters.
What can I feed into the model?
It accepts text and image input, with a 512K context window.
Can I run it on standard serving tools?
vLLM and SGLang support is landing via open PRs.
Send on LINE

AI news that matters for your work, delivered every morning.

Pick your industry and the AI tools you use, and get news related to your work every day.

Free · 30 seconds with Google · unsubscribe anytimeWhat is AIToday? →

Ask AI

Ask AI anything about this article. The AI reads this article, earlier AIToday articles, and Wikipedia, and cites its sources. Q&As are published on this page for other readers too.

Questions and answers are published on this page.

Related Articles

Next articleAleph named exclusive ChatGPT Ads reseller in 44 countries