A smaller parameter count does not necessarily mean a shallower computation: this model runs shared middle layers twice, turning 32 distinct decoder layers into 56 layer executions. That design trades additional computation for fewer stored weights, making long-context multimodal inference more accessible on memory-constrained hardware.
SenseNova-Loop-9B
Reuses decoder-layer weights to increase computation depth at roughly 9B parameters, taking text and images as input and generating text with up to 512K-token context. It targets reasoning, coding, and tool use, though performance on complex long-horizon tasks can be inconsistent.
Introduction
Information
- Websitehuggingface.co
- OrganizationsSenseTime
- Published date2026/11/17
Categories
More Items
A 4B reasoning model with refusal behavior modified using Heretic, built on a student distilled from DeepSeek-V4-Pro. Its model card reports 5 refusals in a 100-prompt test, compared with 99 for the original.
Handles text and image inputs and generates text, with a context window of up to 1 million tokens. Its hybrid sparse and sliding-window attention supports long-context work, while the MoE design activates about 12B parameters per token.
Generates 480p video with two-step autoregressive diffusion, reaching about 45 ms per frame on eight B200 GPUs. Distilled from Wan2.1-14B, it targets interactive generation but needs a multi-GPU setup for that speed.