Blog Network

SenseNova · 2026-11-17 · notable

SenseNova-Loop-9B — an open 9B model that runs its middle layers twice

SenseNova-Loop-9B is an Apache-2.0 multimodal model with 9B parameters and a 512K context. A looped block runs its middle layers twice, and it beats its October preview on all seven benchmarks shown.

Hugging Face model card preview for sensenova/SenseNova-Loop-9B

SenseNova-Loop-9B gets the compute of a bigger model from 9B weights by reusing its middle layers.

Key specs

Context window512K tokens
Terminal bench 2.159.55

What is it?

SenseNova-Loop-9B is the full release of SenseNova's looped open-weight model, out on 17 November 2026 after a preview on 31 October. It reads text and images, writes text, handles up to 512K tokens of context, and ships under Apache-2.0 with an official AWQ 4-bit version.

How does it work?

The network has 32 distinct decoder layers, but layers 4–27 form a block that runs twice, so each forward pass does 56 layer executions — about 14.6B inference-equivalent parameters. Two-stream manifold-constrained hyper-connections (mHC) mix the residual streams, and every group of four layers pairs three linear-attention layers with one full-attention layer.

Why does it matter?

With 4-bit weights and an 8-bit KV cache, SenseNova says multimodal inference at 256K context fits in about 15 GiB of GPU memory. Against the preview, its vendor-reported scores jump on agent tasks: Terminal-Bench 2.1 rises from 40.45 to 59.55 and GAIA from 37.86 to 69.90. The model card also lists weak spots: complex instruction following with many tools, and uneven coding and agent results.

Who is it for?

developers running open models on a single GPU

Try it

pip install "transformers==5.18.0" accelerate safetensors, then load sensenova/SenseNova-Loop-9B with trust_remote_code=True

Sources · 2 outlets

Tags

  • sensenova
  • sensetime
  • open-weights
  • looped-transformer
  • multimodal
  • vision-language
  • long-context
  • apache-2.0
  • hugging-face

← All releases