Successor arrives about two and a half weeks after October preview

China's SenseTime released the weights (trained parameters) of "SenseNova-Loop-9B," an AI model with about 9 billion parameters, on Nov. 17 in "sensenova," its official repository on the AI model-sharing site Hugging Face. According to Hugging Face's public API, the repository was created at 7:16 p.m. Nov. 17 Japan time and last updated at 9:19 p.m. As of 10:22 p.m. that day, it had 0 downloads.

It succeeds "SenseNova-6.8-Loop-9B-Preview," a preview the company released at the end of October (see ORIENS's earlier report for details on the preview). The update history on the model card (the model's description page) says the company released the preview on Oct. 31 and the new model on Nov. 17, each with weights and code. The card describes the new model with the words "officially release."

Downloading the weights requires agreeing to conditions

The license is Apache 2.0. The public API also tags the model "license:apache-2.0."

However, the weight files cannot be downloaded freely. The page says the repository is publicly accessible but that users must accept conditions to access its files and content, asking them to agree to share their Hugging Face information. The public API also shows the model is set to require access requests (gated=auto). The description and specifications can be read without logging in, but running the model requires logging in to Hugging Face and submitting a request.

56 layer computations from 32 layers

According to the model card, the design carries over from the preview. Of its 32 layers, inputs pass through the first 4 layers, then twice in a row through the middle 24 layers using the same weights, and then through the last 4 layers. Each forward pass performs 56 layer computations while storing weights for only 32 layers. The model has about 9 billion parameters, and inference is equivalent to about 14.6 billion parameters, the card says. The public API shows about 9.41 billion weights.

In the middle layers, the model uses "mHC (manifold-constrained hyper-connections)," which carries information along two pathways and learns how each layer reads from and writes to them. For attention (the computation that decides which parts of the context to refer to), every four layers combine three lighter linear-attention layers with one full-attention layer that looks at the whole context. It accepts text and images and responds in text, with a context window of up to 512,000 tokens (512K). The weights are in BF16.

The card also shows experiments that varied the number of passes through the middle layers from one to four. Results across four tasks were best at three or four passes, but both the preview and the new model use two passes by default to balance inference efficiency, the card says.

About 15 GiB of GPU memory with 4-bit compression

According to the card, compressing the weights to 4 bits brings their size to about 4.9 GB. With 4-bit weights and an 8-bit KV cache (memory that holds the context during processing), the company estimates that inference including images at a 256,000-token (256K) context requires about 15 GiB of GPU memory. It says it also released a 4-bit AWQ version it prepared on launch day.

Guidance for inference software is adapted from the preview's instructions. For vLLM and SGLang, the card directs users to versions from change proposals (pull requests) that add support for the model. It says inference software versions and settings for the new model will be updated later.

Weaknesses in instruction-heavy and long tasks

The card lists three "known limitations." When given many complex tool definitions or multiple constraints, the model may misunderstand or miss requirements. Its performance on programming and general agent tasks (where AI carries out steps on its own) is inconsistent, and its ability to handle complex, long tasks still needs improvement, the card says. Websites with complex 3D scenes or many animations may not be completed in one turn, and some examples needed two to three rounds of refinement based on user feedback. The company said it plans to use the design in larger models.

Scores up from the preview; figures self-reported by the company

According to the model card, the company compared the new model with the preview and other models on seven agent, programming and data analysis benchmarks. The company says the new model scores higher than the preview on all seven benchmarks shown. AutomationBench went from 14.33 to 36.67, Terminal-Bench 2.1 from 40.45 to 59.55, and GAIA from 37.86 to 69.90. BIRD went from 71.38 to 72.34, a difference of 0.96 points.

The card also shows a comparison with "Qwen3.8-27B." On AutomationBench both scored 36.67, and on AIDABench the new model scored 65.31 and Qwen3.8-27B 65.01. On NL2Repo, the new model scored 41.36 and Qwen3.8-27B 44.60.

All figures are the company's own evaluations on the model card, and no third-party verification was found. The card also shows the results for all seven benchmarks in a table image, but the image could not be retrieved, so the table below lists only the figures for the four benchmarks written in the card's text.

Scores of the preview (SenseNova-6.8-Loop-9B-Preview) and the new model. SenseTime's own evaluation figures written in the model card's text
  • Preview
  • SenseNova-Loop-9B
  • AutomationBench14.33 → 36.67
  • Terminal-Bench 2.140.45 → 59.55
  • GAIA37.86 → 69.90
  • BIRD71.38 → 72.34