SenseNova open-sources Loop-9B: 9B multimodal model with 56 weight-shared layer executions, 256K context
AdinaYakup · x · 2026-11-02
SenseNova released SenseNova-Loop-9B-Preview, a multimodal model built on a looped, weight-sharing architecture, now up on Hugging Face.
Key details:
- 9B parameters, Apache 2.0 license
- Executes 56 layer passes via weight sharing, trading parameter count for effective depth
- Up to 256K context
- Hybrid attention plus a two-stream mHC design
- Chat template handles multi-image and multi-video inputs
A preview-stage open-source take on the looped-depth approach worth tracking for future evals.
Related event: SenseTime Open-Sources SenseNova-Loop-9B Multimodal Model(2 posts)→
More from Multimodal
- Suno Studio adds track hiding, grid controls, and semitone/cents fine pitch editing — suno · 2026-11-03
- Suno Studio update adds clip settings sidebar, fine pitch tuning and auto-stretch controls — suno · 2026-11-03
- Lockstep video recast beats rivals at 3.3x lower cost in hands-on test — alifcoder · 2026-11-03
- Hume Launches Audio Decision API Answering Voice Questions in One Call — EXM7777 · 2026-11-03
- Restoring a 1970 Super 8 reel with Topaz Astra in Adobe Firefly: a 4-step upscale workflow — rufusd · 2026-11-03
- Lockstep motion transfer model lets filmmakers recast scenes from a reference image — azed_ai · 2026-11-03