SenseNova-Loop-9B runs its middle layers twice, openly

What's new

SenseTime published SenseNova-Loop-9B-Preview as open weights on Hugging Face. It accepts text and images and outputs text. It targets long-context and agentic work: document understanding, coding and tool calls. Training context reaches 262,144 tokens.

How the loop works

The model has 32 distinct decoder layers. After layers 0 to 3, the block of layers 4 to 27 runs twice in a row with shared weights, followed by the last 4 layers. One forward pass therefore performs 56 layer computations while storing only 32 layers of weights. SenseTime describes this as about 9B stored and about 14.6B inference parameters.

• Two-stream mHC: two residual streams run through the looped section, mixed by learned mappings with Sinkhorn normalization.
• Hybrid attention: each four-layer group has three linear-attention layers and one full-attention layer.
• Acceleration: MTP is available, while Dspark and Dflash2 are still in training.

Benchmarks

Source: SenseTime. SenseNova-Loop-9B-Preview's own results on eight benchmarks, next to Qwen3.5/3.6 and other models.

In SenseTime's chart, Loop-9B scores 71.38 on BIRD-Dev versus 69.84 for Qwen3.5-27B. On WorkBuddy Bench it reaches 66.89, against 61.98 for the best comparison model. It also leads on AutomationBench and the in-house RaccoonPPTBench. It does not win everywhere: on Terminal Bench 2.1 it scores 40.45, behind Qwen3.6-27B at 47.19, and Qwen3.5-27B stays ahead on τ³-Banking. All figures are self-reported, and the technical report is still pending (echina.news).

Why it matters

Looped transformers are an old idea, but usable open multimodal models built on them are rare. If the numbers hold up in independent testing, this is a cheaper path: more compute from the same memory. Running it needs Transformers 5.18.0 with trust_remote_code, and compatible inference engines must reproduce the loop order.