HaiAI123

Curated Global AI Tools Directory

Industry update
Flat

SenseNova-Loop-9B Official Release Open-Sourced with Domestic GPU Support

11/17/2026 — 11/18, 01:20·3 sources (all)·3 reports (all)

Key takeaways

  • SenseNova has open-sourced the official SenseNova-Loop-9B, a recurrent Transformer with 32 stored layers, 24 of which run twice, yielding 56…

Background & analysis

On the evening of November 17, 2026, SenseTime unveiled the SenseNova Loop-9B model. The release continues the parameter-efficiency line of its Preview version, reusing layer parameters through a Loop structure and applying manifold-constrained hyper-connections (mHC) to extend computation depth at roughly a 9B parameter scale, targeting complex reasoning, coding, tool calling, and multimodal workloads.

Technically, the model groups layers 4 through 27 into a shared-weight recurrent module. By executing the shared intermediate layers twice, it expands 32 physical decoding layers into 56 layer computations, with total parameters of about 9B. It accepts text and image inputs, produces text outputs, and supports a context window of up to 512K tokens. According to the company, after 4bit quantization the weights measure roughly 4.9GB, keeping the deployment memory requirement within 10GB, which widens the range of hardware capable of running the model.

SenseTime also states that SenseNova Loop-9B performs competitively with Qwen3.8-27B on agent and data-analysis benchmarks. It scores 41.36 on NL2Repo, close to the 44.60 recorded by Qwen3.8-27B, and 65.31 on AIDABench, above the 44.60 baseline counterpart's 65.01. For now, the available record consists of this launch and the company's own description, with no further developments reported.

AI-generated from 1 reports

Latest turnSenseTime has released SenseNova Loop 9B, an open-weight multimodal model that handles images, code, data analysis and tool use, and can run on local devices. In the company's experiments, keeping the 9B weights fixed while raising training and inference loops from one to 3-4 pushed scores on LiveCodeBench V6 and office data-analysis tasks up 6.4%-16.9%. The model also supports Chinese GPUs from Cambricon, Moore Threads and Hygon.

updream
24-hour heatHeat index 242 · peak 284 · 5h ago
24 hours agonow

Reports on this story headlines open the original

Today
  1. SenseTime has released SenseNova Loop 9B, an open-weight multimodal model that handles images, code, data analysis and tool use, and can run on local devices. In the company's experiments, keeping the 9B weights fixed while raising training and inference loops from one to 3-4 pushed scores on LiveCodeBench V6 and office data-analysis tasks up 6.4%-16.9%. The model also supports Chinese GPUs from Cambricon, Moore Threads and Hygon.

    V2EXAI score 56
Yesterday
  1. SenseNova has released SenseNova-Loop-9B, a looped Transformer that stores 32 layers and runs 24 of them twice, doing 56 layers of work per forward pass at roughly 9B parameters under Apache 2.0. Its README reports gains over the Preview on all seven benchmarks, including AutomationBench rising from 14.33 to 36.67 and GAIA from 37.86 to 69.90.

    Reddit · LocalLLaMAAI score 58
  2. SenseTime has released SenseNova Loop-9B, which reuses a shared block of middle layers to turn 32 physical decoder layers into 56 layer computations, packing roughly 9B total parameters, text and image input, and up to 512K tokens of context. SenseTime says 4-bit weights take about 4.9GB and deployment needs under 10GB of VRAM, with agent and data-analysis benchmark scores close to Qwen3.8-27B.

    IT之家AI score 58

Other stories people are talking about

How is heat calculated?About the method

Heat counts how many independent sources covered a story in the last 48 hours: one source counts once no matter how many posts it published, decaying with a 24-hour half-life. What ranks first is what many people are talking about.

This page aggregates public feeds. Headlines and summaries are machine-organized and remain the property of the original authors; verify important facts at the source. If you believe a headline or summary infringes your rights, email the contact address in the footer with the page URL and basis of your claim, and we will remove or replace it after review.

Surge
Discussion rising fast
New
First report within 6 hours
Rising
Still gathering discussion

Back to AI industry updates →