DeepSeek's MoE model with 552B total parameters and 8-16B active per token. 1M-token context, MIT license: free commercial use. Built for online serving, too large for local use on a single PC.
SenseNova-Loop-9B
Released on Nov 17, 2026
Multimodal model of roughly 9B with a looped architecture: it reuses the weights of its 24 middle layers twice, extending 32 physical layers to 56 executions, with two-stream mHC hyper-connections and a context of up to 512K tokens. With 4-bit weights and an 8-bit KV cache it needs roughly 15 GiB of VRAM for a 256K context. On AIDABench it beats Qwen3.8-27B (65.31 vs 65.01); GAIA 69.90.
Features
- Architettura a loop (32β56 esecuzioni)
- Contesto 512K token
- Multimodale: testo + immagine
- ~15 GiB VRAM a 256K (pesi 4-bit, KV 8-bit)
- Quantizzazione AWQ 4-bit ufficiale
Version history
- SenseNova-Loop-9BLatest
Release ufficiale: blocco a loop con mHC a due flussi, contesto esteso a 512K, quantizzazione AWQ 4-bit ufficiale; GAIA 69,90 e Terminal-Bench 2.1 59,55.
- SenseNova-Loop-9B-Preview
Primo rilascio di pesi e codice: architettura a loop con mHC, ~9,4B parametri, orientata all'efficienza parametrica.
Tags
Useful links
Discover SenseNova-Loop-9B
Visit the official site to download, learn more and start using SenseNova-Loop-9B.
Related resources β Models
The most advanced open image model: excellent prompt adherence, solid anatomy, full fine-tuning. The standard for serious work.
Very few steps, high quality, and strong text rendering. The distilled FLUX variant, built for speed.

