# SenseNova-Loop-9B

> SenseNova's open-weight looped multimodal model, released 17 November 2026 — about 9B parameters whose middle 24 layers run twice per forward pass, 512K context, text and image in, under Apache 2.0.

- Maker: SenseNova
- Line: SenseNova-Loop (looped open weights)
- Canonical page: https://ai-tldr.dev/models/sensenova-loop-9b/
- Updated: 2026-11-17

## In short

SenseNova-Loop-9B is an open-weight multimodal language model that SenseNova, SenseTime's foundation-model family, released on Hugging Face and GitHub on 17 November 2026, following SenseNova-Loop-9B-Preview on 31 October 2026. It reads text and images, writes text, handles up to 512K tokens of context and supports a thinking mode. It has a 512K-token context window. You can use it through Hugging Face (weights).

## Facts

| Fact | Value | Source |
| --- | --- | --- |
| Context window | 512K tokens | [1] |
| Licence | Apache 2.0 · open weights | [1] |
| Released | 17 November 2026 | [2] |
| Parameters | ≈9B total (≈14.6B inference-equivalent with the looped block) | [1] |

## Sources

1. [sensenova/SenseNova-Loop-9B on Hugging Face](https://huggingface.co/sensenova/SenseNova-Loop-9B)
2. [OpenSenseNova/SenseNova-Loop on GitHub](https://github.com/OpenSenseNova/SenseNova-Loop)

## Overview

SenseNova-Loop-9B is an open-weight multimodal language model that SenseNova, SenseTime's foundation-model family, released on Hugging Face and GitHub on 17 November 2026, following SenseNova-Loop-9B-Preview on 31 October 2026. It reads text and images, writes text, handles up to 512K tokens of context and supports a thinking mode. The weights and code are under the Apache License 2.0, and SenseNova published an official AWQ 4-bit version on launch day.

Its defining idea is a looped decoder. The model stores 32 distinct decoder layers, about 9B parameters, but layers 4–27 form a block that runs twice with the same weights: inputs pass through layers 0–3, through the block twice, then through layers 28–31 and the output head. Each forward pass therefore performs 56 layer executions, which SenseNova describes as about 14.6B inference-equivalent parameters. Inside the loop, two-stream manifold-constrained hyper-connections (mHC) carry two residual streams and mix them through a Sinkhorn-normalised map, and every group of four layers pairs three linear-attention layers with one full-attention layer.

SenseNova's loop-scaling experiments, which vary the loop count with training data held constant, found the best results at three or four loops; the release uses two by default to keep inference fast. Because only 9B distinct parameters are stored, SenseNova estimates that with 4-bit weights (about 4.9 GB) and an 8-bit KV cache, multimodal inference at 256K context needs about 15 GiB of GPU memory. The README lists its known weak spots: complex instruction following across many tool definitions, uneven coding and general agent results, and complex 3D or animated websites that can take two to three rounds of refinement.

## Benchmarks

- GAIA: 69.9% ([source](https://github.com/OpenSenseNova/SenseNova-Loop))
- BIRD: 72.34% ([source](https://github.com/OpenSenseNova/SenseNova-Loop))
- AIDABench: 65.31% ([source](https://github.com/OpenSenseNova/SenseNova-Loop))
- Terminal-Bench 2.1: 59.55% ([source](https://github.com/OpenSenseNova/SenseNova-Loop))
- NL2Repo: 41.36% ([source](https://github.com/OpenSenseNova/SenseNova-Loop))
- AutomationBench: 36.67% ([source](https://github.com/OpenSenseNova/SenseNova-Loop))
- τ³-Banking: 30.93% ([source](https://github.com/OpenSenseNova/SenseNova-Loop))

## Strengths

- Looped architecture: 32 stored decoder layers run as 56 layer executions per forward pass, about 14.6B inference-equivalent parameters from 9B weights
- 512K-token context with text and image input
- Low-memory deployment: about 15 GiB of GPU memory for multimodal inference at 256K context with 4-bit weights and an 8-bit KV cache, by SenseNova's estimate
- Vendor-reported gains over the Preview on all seven benchmarks shown — Terminal-Bench 2.1 from 40.45 to 59.55 and GAIA from 37.86 to 69.90
- Apache 2.0 weights and code, an official AWQ 4-bit version, and MTP speculative decoding support
- Deployment examples for Transformers, LightLLM, SGLang and vLLM

## Best for

- Reach for it to run an open multimodal model with long context on a single consumer or workstation GPU
- Reach for it for agentic, tool-use and data-analysis workloads where a small model's memory footprint matters
- Reach for it to research looped transformers and parameter sharing with an openly licensed checkpoint and modelling code
- Reach for it as a base to fine-tune with the SenseNova-Loop examples in OpenSenseNova/torchtitan

## How to access

- [Hugging Face (weights)](https://huggingface.co/sensenova/SenseNova-Loop-9B) — model id `sensenova/SenseNova-Loop-9B`

## SenseNova-Loop (looped open weights) versions

- SenseNova-Loop-9B (17 November 2026): Full release: same looped architecture and two-stream mHC as the Preview, with higher agentic, coding and data-analysis scores and an official AWQ 4-bit version. — this page
- SenseNova-Loop-9B-Preview (31 October 2026): First public release of the looped 9B model's weights and code.

## FAQ

### What is SenseNova-Loop-9B?

SenseNova-Loop-9B is an open-weight multimodal model from SenseNova, SenseTime's foundation-model family, released on 17 November 2026. It has about 9B parameters, reads text and images, writes text and supports up to 512K tokens of context.

### What does 'looped' mean in SenseNova-Loop-9B?

Its middle decoder layers (layers 4–27) form a block that runs twice per forward pass with the same weights. The model stores 32 layers but performs 56 layer executions, which SenseNova describes as about 14.6B inference-equivalent parameters from 9B stored ones.

### What licence is SenseNova-Loop-9B under?

The model weights and code are released under the Apache License 2.0.

### How much GPU memory does SenseNova-Loop-9B need?

SenseNova estimates that with 4-bit weights (about 4.9 GB) the model needs about 10 GiB at 64K context, 12 GiB at 128K and 15 GiB at 256K with an 8-bit KV cache, or about 9, 10 and 12 GiB with a 4-bit KV cache.

### How do I run SenseNova-Loop-9B?

Install a CUDA build of PyTorch, then `python -m pip install "transformers==5.18.0" accelerate safetensors`, and load `sensenova/SenseNova-Loop-9B` with `AutoModelForCausalLM.from_pretrained(..., trust_remote_code=True)`. The README also gives serving examples for LightLLM and for SGLang and vLLM branches that include the SenseNova Loop implementation.

## Links

- [sensenova/SenseNova-Loop-9B on Hugging Face](https://huggingface.co/sensenova/SenseNova-Loop-9B)
- [OpenSenseNova/SenseNova-Loop on GitHub](https://github.com/OpenSenseNova/SenseNova-Loop)
- [sensenova/SenseNova-Loop-9B-Preview on Hugging Face](https://huggingface.co/sensenova/SenseNova-Loop-9B-Preview)
- [SenseNova platform](https://platform.sensenova.cn/)

Source: AI/TLDR LLM registry, https://ai-tldr.dev/models/sensenova-loop-9b/
