Model Card | SenseNova Loop-9B-Preview

🌏 中文版

Model Info

Highlights

• Looped decoder: the input first passes through layers 0-3, then the block spanning layers 4-27 is applied twice in sequence (B(B(h))), before layers 28-31 and the output head produce the result. Inference performs 56 decoder-layer computations while storing only 32 sets of weights.
• Two-stream mHC (Manifold-Constrained Hyper-Connections): inside the looped section, two residual streams carry information through learned mappings, with Sinkhorn normalization (20 iterations) constraining the stream-mixing map. An official ablation shows adding mHC alone lifts GPQA Diamond by 4.42 points and TeamBench by 13.34 points.
• Hybrid attention: each group of four layers interleaves three linear-attention layers with one full-attention layer, and the loop boundaries align with complete four-layer groups so the attention pattern stays consistent across both passes.
• At 9B parameters, it beats the 3x-larger Qwen3.5-27B on WorkBuddy Bench, AutomationBench, and RaccoonPPT Bench (confirmed by a third-party account, TeksEdge, who verified the numbers against SenseTime's own launch chart).

Benchmarks

⚠️ The three task scores above come from SenseTime's official launch chart, transcribed after verification by third-party X account TeksEdge — not an official transcript, and not yet independently reproduced. The official chart also lists BIRD-Dev, AIDABench, Terminal Bench 2.1, and T³-Banking; on the latter two, the chart shows Loop-9B trailing Qwen3.5-27B, but the exact scores have no published transcript, so they're excluded from the table.

vs. Predecessor and Competitors

SenseNova Loop is the first public release in this architecture branch from SenseTime, so there's no strict "predecessor" to compare against. The official comparisons are against same-size Qwen3.5-9B (where a third party found Loop-9B winning 7 of 8 charted benchmarks and tying the eighth) and the deliberately harder target, Qwen3.5-27B (3x the parameters).

Against Qwen3.5-27B, Loop-9B wins on office and automation-style tasks (WorkBuddy, AutomationBench, RaccoonPPT all lead), but trails on Terminal Bench 2.1 (terminal operations) and T³-Banking (financial workflow simulation). That split suggests the gains from "looped compute depth" aren't distributed evenly across task types — on tasks requiring rigorous multi-step tool-chain execution and long transaction sequences, it still falls short of simply scaling parameters to 27B.

There's no pricing strategy to compare: Loop-9B currently ships as open weights only, with no official API, which puts it on a completely different business model from the cloud models announced around the same time, like Gemini 4 Carbon or Kimi K3.1 Pro. What it's selling is architectural efficiency — trading stored parameters for compute depth — not price competition.

What This Means for Agent Development

The most direct implication of the Looped Transformer architecture for agent development is trading a deeper compute path for reasoning ability without growing the model's storage footprint — 56 layer computations fit in the memory footprint of 32 stored weight sets, a different trade-off from the usual "more parameters, more intelligence" path.

• If you're running agents on consumer or edge GPUs (like a 16GB RTX card): this model line is worth watching. It achieves near-27B performance on office and automation tasks at a 9B memory footprint, which is a better cost-to-capability ratio in resource-constrained settings.
• If you're building for multi-step tool-chain execution or long transaction reasoning (terminal-operation automation, financial workflows): SenseTime's own chart shows Loop-9B trailing the same-family 27B model on Terminal Bench 2.1 and T³-Banking, so it isn't a drop-in replacement for your current tool-chain agent yet.
• Not a fit: scenarios that need a stable public API, SLA guarantees, or off-the-shelf pricing — there's only open weights for now, no official hosted service, and self-hosting means dealing with custom inference-framework implementations (LightLLM or SGLang, loaded with trust_remote_code=True for the custom model code).

Today's Takeaway

This one points to a direction different from "more parameters, more intelligence": re-running the same weights to deepen the effective compute path, instead of storing more parameters. For agent development, that hints model size and reasoning depth might eventually decouple, which would change how deployment cost gets calculated. But SenseTime's own ablation chart also shows the gains aren't evenly distributed — it trails the same-family 27B model on Terminal Bench and financial workflows, which suggests "trading compute depth for intelligence" is still a targeted optimization, not yet a general-purpose solution.

References

• SenseNova-Loop-9B-Preview model card (Hugging Face)
• OpenSenseNova/SenseNova-Loop GitHub source and README
• SenseTime on X: Introducing SenseNova Loop-9B-Preview
• David Hendrickson (@TeksEdge) on X: third-party verification of SenseNova Loop-9B-Preview's numbers
• SenseTime open-sources SenseNova Loop-9B-Preview: exploring compute depth to unlock model potential (Sina Finance, in Mandarin)