huggingface.co web signal

SenseNova Open-Sources Loop-9B Model With 512K Context

Open Source Multimodal ai-business

TL;DR

  • SenseNova-Loop-9B stores about 9 billion parameters but executes 56 decoder-layer passes per forward step, giving roughly 14.6 billion inference-equivalent compute.
  • The release ships under Apache 2.0 with a 512K context window, text-plus-image input, and an official AWQ 4-bit variant that fits about 10 GiB at 64K context.
  • On AutomationBench it ties Qwen3.8-27B at 36.67 and edges it on AIDABench 65.31 to 65.01, but trails on NL2Repo at 41.36 against 44.60.

SenseNova released SenseNova-Loop-9B under Apache 2.0 on November 17, with roughly 9 billion stored parameters that behave like 14.6 billion at inference time. The mechanism is a looped decoder: 32 distinct layers, 56 layer executions per forward pass.

The model card lays out the shape exactly. 'Layers 0–3 → Looped Block (layers 4–27, executed twice) → Layers 28–31,' with shared weights across both passes of the middle block. Attention alternates in repeating four-layer groups, 'Three linear-attention layers followed by one full-attention layer per group,' and two residual streams route information through 20 Sinkhorn iterations of what SenseNova calls Manifold-Constrained Hyper-Connections.

On benchmarks the 9B pulls level with a 27B peer in places. AutomationBench scores land at 36.67 for both SenseNova-Loop-9B and Qwen3.8-27B. AIDABench goes to the smaller model, 65.31 to 65.01. NL2Repo goes the other way, 41.36 to 44.60. The preview-to-release jumps are large on some axes: Terminal-Bench 2.1 climbed from 40.45 to 59.55, GAIA from 37.86 to 69.90, BIRD nudged from 71.38 to 72.34.

Context is 512K tokens. Input is text and images, output is text. In 4-bit AWQ with 8-bit KV cache the model needs about 10 GiB at 64K context and 15 GiB at 256K, per the memory table on the card. vLLM support is landing via PR #64832 and SGLang via PR #47281; neither PR is listed as merged.

SenseNova is explicit about where the model falls over. 'Complex instruction following may struggle with multiple constraints or complex tool definitions,' and complex web generation 'requires multiple refinement rounds for 3D/animated content.' The drop lands on the same day as MBZUAI's six-model K2-Horizon fleet, both into an open-weight stack that AI Weekly has tracked across 384 stories in the past 90 days.