SenseNova-Loop-9B Leaves Preview With Its GAIA Score Nearly Doubled and the Same Two-Pass Layer Design

SenseTime's SenseNova team released the official SenseNova-Loop-9B on Hugging Face on November 17, 2026, under Apache 2.0 with gated downloads. The 9B multimodal model still runs 24 shared layers twice, now posts large self-reported gains over the October preview on agentic and coding benchmarks, and ships with a loop-depth study and low-memory deployment guidance.

A spiral staircase seen from above inside a stone tower, its middle flights looping back on themselves as a lone walker climbs, lit in fuchsia and dove grey.
A spiral staircase seen from above inside a stone tower, its middle flights looping back on themselves as a lone walker climbs, lit in fuchsia and dove grey.

SenseTime’s experiment in reusing layers is out of preview. On November 17, 2026, the company’s SenseNova team published SenseNova-Loop-9B in its official Hugging Face organisation, releasing weights and code two and a half weeks after the October 31 preview. The model card calls it the official release and says it builds on community feedback about use, inference deployment and ecosystem integration.

The repository is tagged Apache 2.0, but the files are gated. Hugging Face shows that the repository is public while the weights require users to log in and agree to share their account information before downloading. The model card and its specifications can be read without logging in.

Same trick, more depth per parameter

The design carries over from the preview. SenseNova-Loop-9B stores 32 distinct decoder layers. Input goes through layers 0 to 3, then through the same block of layers 4 to 27 twice in a row using shared weights, and then through layers 28 to 31 and the output head. Each forward pass therefore executes 56 layer computations while holding weights for only 32. The card puts the model at about 9 billion parameters, equivalent to roughly 14.6 billion at inference; Hugging Face’s API counts about 9.41 billion weights.

Two other pieces of the architecture matter. The decoder uses hybrid attention, with three linear-attention layers followed by one full-attention layer in every group of four, and the loop boundaries line up with those groups so both passes keep the same pattern. Inside the looped block, two-stream manifold-constrained hyper-connections (mHC) carry hidden states in two parallel residual streams, with learned maps deciding how each layer reads, mixes and writes them. The mixing is constrained with 20 Sinkhorn normalisation iterations, which the card notes are separate from the two passes through the loop.

The model takes text and images and outputs text, and its training extended context to 512K tokens. Hidden size is 4,096, the feed-forward dimension 12,288 and the vocabulary 248,320 tokens.

What changed from the preview

The headline difference is performance on agent-style work. SenseTime says the release beats the preview on all seven benchmarks it shows. AutomationBench rises from 14.33 to 36.67, Terminal-Bench 2.1 from 40.45 to 59.55, and GAIA from 37.86 to 69.90. The company also compares the 9B model with Qwen3.8-27B, a model with three times as many parameters. It reports a tie at 36.67 on AutomationBench, 65.31 against 65.01 on AIDABench and 41.36 against 44.60 on NL2Repo. On the BIRD text-to-SQL test the gain is small, from 71.38 to 72.34. All figures are SenseTime’s own evaluations.

Does more looping help?

The card includes a study of loop depth. Holding training data and settings constant, SenseTime compared one to four passes through the shared block. Performance improved across four tasks, with the best results at three or four loops. AIDABench-QA and Workbuddy-office peaked at three loops, up 6.42% and 16.93% over a single pass. LiveCodeBench V6 and AIDABench data visualisation peaked at four, up 8.10% and 8.55%.

Both the preview and the release default to two loops to keep inference efficient. Going from two to four loops added 1.04 points on LiveCodeBench V6 and 1.35 points on AIDABench data visualisation. SenseTime says it will keep exploring deeper looping and “Dynamic Loop” architectures that adapt the number of passes to the difficulty of each input.

About 4.9 GB of 4-bit weights and pull-request builds of SGLang and vLLM

Because only about 9 billion distinct parameters are stored, the memory footprint is small for the depth. With 4-bit weights the card says the weights take about 4.9 GB. Its estimates for minimum GPU memory with an 8-bit KV cache are about 10 GiB at 64K context, 12 GiB at 128K and 15 GiB at 256K, with a 4-bit cache trimming each by 1 to 3 GiB. The card says an official AWQ 4-bit version is available on launch day.

Loading through Transformers 5.18.0 requires trust_remote_code=True for the custom model code, and thinking mode can be toggled in the chat template. Serving examples cover LightLLM, plus SGLang and vLLM by checking out specific pull requests (#47281 and #64832) rather than released versions. The deployment commands are adapted from the preview, and the team says release-specific engine settings will follow. Fine-tuning examples live in the OpenSenseNova torchtitan fork, and multi-token-prediction speculative decoding drafts are covered in the SenseSpec repository.

SenseTime lists known limitations. The model can miss requirements when given many complex tool definitions or constraints, performs inconsistently on coding and general agent tasks, and may need two or three rounds of feedback to finish complex 3D web pages.

Topics
  • #SenseTime
  • #SenseNova
  • #Open Weights
  • #Looped Transformer
  • #Apache 2.0
Raj M

Author

Raj M

Contributor

AI Systems Architect is a seasoned technology leader with over 15 years of experience in the IT industry working with Fortune 500 companies. With a solid foundation in multi-agent systems, open-source LLM infrastructure, and enterprise deployment, he excels at building scalable production-grade AI platforms.