You need to agree to share your Hugging Face information to access this model

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

Log in or Sign Up to review the conditions and access this model content.

SenseNova-Loop-9B

English | 简体中文

SenseNova-Loop-9B on Hugging Face SenseNova-Loop-9B-Preview on Hugging Face SenseNova-Loop-9B-Preview on ModelScope GitHub repository Apache 2.0 License sensenova_loop architecture Join Discord

🌐 Architecture

这张图是SenseNova-Loop-9B的整体架构示意图,整体以淡紫色调呈现。图的上部展示模型的宏观层级流转:先经过前置循环的A层,随后进入标注为Looped block的循环模块,该模块包含1-st Looped Pass等多轮循环传递,每轮均包含多个TBK单元,对应模型中部的B层,之后进入后置循环的最后C层,最终接入LM Head得到输出。图的下部为mHC模块的内部细节,展示了特征通过Transform分支的Attention/FFN计算、结合双随机矩阵的残差混合多流操作,完成特征的迭代更新,直观呈现该模型的核心特殊架构设计。

📣 News

🔥 2026-11-17: We release SenseNova-Loop-9B, with model weights and code available to the community.

🔥 2026-10-31: We release SenseNova-Loop-9B-Preview, with model weights and code available to the community.

🚀 SenseNova-Loop-9B: Deeper Computation with Fewer Parameters

We officially release SenseNova-Loop-9B. Building on the Preview's exploration of parameter efficiency, the model reuses layer weights through a looped architecture and combines them with manifold-constrained hyper-connections (mHC) to expand computational depth within a compact parameter budget. It is designed for complex reasoning, coding, tool use, and multimodal tasks.

Since the Preview release, we have received community interest and feedback on model use, inference deployment, and ecosystem integration. The official release retains the looped architecture and two-stream mHC, further improving performance on agentic, coding, and data analysis tasks while continuing to explore the capabilities of looped computation.

SenseNova-Loop-9B has approximately 9B parameters. By repeatedly executing shared middle layers, it extends 32 physical decoder layers into 56 layer executions. The model accepts text and images and generates text, supporting long-context understanding and multi-step tasks. In agentic and data analysis evaluations, the model is competitive with Qwen3.8-27B: it scores 41.36 versus 44.60 on NL2Repo and 65.31 versus 65.01 on AIDABench.

Model Highlights

  • Shared looped block: 24 decoder layers are executed twice in sequence, producing 56 layer computations from 32 distinct layers per forward pass.
  • Two-stream mHC: learned mappings route and combine information between two residual streams within the looped block.
  • Hybrid attention: the decoder interleaves linear and full attention in a repeating four-layer pattern.
  • Long-context multimodal training: the training recipe includes text, images, code, and tool-use trajectories, with the context length extended to 512K tokens.
  • Low-memory deployment: With 4-bit weights and an 8-bit KV cache, multimodal inference at a 256K context length requires approximately 15 GiB of GPU memory.
  • Official quantized release: An official AWQ 4-bit version is available on launch day for easier deployment.

🎬 Showcase

3D

Cloudside Market: rotating 3D scene

SVG + Web

ORBIT: an interactive space exploration website

Poster

Distortion FM: an animated music festival poster

Web

TYPE RELAY: an interactive typography exhibition

Design

Beyond the Coastline: a photo editorial website

📊 Model Performance: From Preview to Release

We evaluate SenseNova-Loop-9B on agentic, coding, and data analysis tasks, comparing the release with the Preview and other models.

Benchmark comparison of SenseNova-Loop-9B, the Preview, and other models

The release outperforms the Preview on all seven benchmarks shown. 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.

Compared with Qwen3.8-27B, the release matches its 36.67 on AutomationBench and scores 65.31 versus 65.01 on AIDABench. On NL2Repo, it scores 41.36 versus 44.60. Its BIRD score of 72.34 is 0.96 percentage points above the Preview's 71.38.

🧠 Model Architecture

SenseNova-Loop-9B is built on a 32-layer multimodal decoder combining shared-weight looped computation, hybrid attention, and two-stream manifold-constrained hyper-connections (mHC). It extends computation depth by reusing middle-layer parameters, accepts text and images, generates text, and supports a context length of up to 512K tokens.

Model Specifications

Property SenseNova-Loop-9B
Architecture Autoregressive multimodal decoder; looped block with two-stream mHC
Parameter scale Approximately 9B total parameters; approximately 14.6B inference-equivalent parameters
Input / output Text and images / text
Distinct decoder layers 32
Looped layer range / passes [4, 28) / 2, with shared weights
Decoder layer executions 56 per forward pass
mHC residual streams 2
Context length 512K tokens
Weight precision BF16 backbone; FP32 mHC routing parameters

Looped Decoder

Inputs first pass through layers 0–3, then through the same block of layers 4–27 twice, and finally through layers 28–31 and the output head. If B denotes the looped block and h its input, the looped computation can be written as B(B(h)).

During inference, each forward pass performs 56 decoder-layer computations while storing only 32 sets of decoder-layer weights. The looped block shares the same weights across both passes, with the output of the first pass serving as input to the second.

Hybrid Attention

The decoder interleaves linear attention and full attention. Each group of four layers contains three linear-attention layers followed by one full-attention layer. The loop boundaries align with complete four-layer groups, so both passes preserve the same attention pattern.

Two-Stream mHC

Within the looped block, mHC uses two residual streams to carry information around decoder-layer computations. Learned mappings determine how each layer reads from the streams, mixes information between them, and writes its output back to the residual state. The stream-mixing map is constrained with Sinkhorn normalization, using 20 iterations in the model configuration.

The two streams can be viewed as two sets of hidden states carried in parallel. A looped layer reads from the streams, applies attention or feed-forward computation, and writes its output back. Repeated row and column normalization constrains how information is mixed between the streams. These 20 normalization iterations are internal to mHC and are separate from the two passes through the looped block.

The loop range, number of passes, and mHC routing parameters together define the checkpoint's execution configuration. A compatible inference implementation must load these parameters and preserve the same computation order.

Configuration Value
Distinct decoder layers / executions 32 / 56
Looped layer range / passes [4, 28) / 2
Hidden dimension 4,096
Feed-forward dimension 12,288
Vocabulary size 248,320
Full-attention heads / KV heads 16 / 4
mHC residual streams / Sinkhorn iterations 2 / 20

🔬 Loop Scaling: Exploring the Benefits of Looped Computation

Loop Scaling: matched training and inference loops across four benchmarks

Looped Transformers show promising performance scaling, offering another way to expand model capability through repeated computation. Using the original 9B model with a standard single forward pass (Loop 1) as the baseline, we vary the number of loops while holding the training data and other settings constant. Overall performance improves across the four tasks, with the best results at three or four loops. AIDABench-QA and Workbuddy-office peak at three loops, improving by 6.42% and 16.93% relative to the baseline; LiveCodeBench V6 and AIDABench data visualization peak at four loops, improving by 8.10% and 8.55%.

To balance inference efficiency, both the Preview and the release use two loops by default. Deeper loops nevertheless provide further gains: moving from two to four loops improves LiveCodeBench V6 by 1.04 percentage points and AIDABench data visualization by 1.35 percentage points. These experiments cover one to four loops, leaving the potential of deeper looping open for further study.

Different tasks respond differently to loop depth. We will continue exploring deeper computation and Dynamic Loop architectures that adapt the number of loops to each input, allocating computation according to task complexity.

💾 Low-Memory Deployment

The shared looped architecture uses approximately 9B distinct parameters. With 4-bit weight quantization, the model weights occupy approximately 4.9 GB. Using an 8-bit or 4-bit KV cache can further reduce memory requirements for long-context inference. The estimated minimum GPU memory requirements for deployment are:

Context length 8-bit KV cache 4-bit KV cache
64K ~10 GiB ~9 GiB
128K ~12 GiB ~10 GiB
256K ~15 GiB ~12 GiB

Get Started

Use Transformers to load the model from a Hugging Face repository or a local model directory. The examples below cover single-GPU text generation and serving with inference engines.

Requirements

The checked Preview environment uses Transformers 5.18.0, PyTorch 2.11.0, and CUDA 13.0. The model directory must include the configuration, tokenizer, chat template, custom model code, and all weight shards.

First install a CUDA-compatible build of PyTorch, then install the remaining dependencies:

python -m pip install "transformers==5.18.0" accelerate safetensors

Text Generation with Transformers

Set model_id to the release model directory or repository ID, then run:

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "/path/to/SenseNova-Loop-9B"
device = "cuda:0"

tokenizer = AutoTokenizer.from_pretrained(model_id, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    trust_remote_code=True,
    dtype=torch.bfloat16,
    device_map={"": device},
).eval()

messages = [{"role": "user", "content": "What is 13 times 7?"}]
prompt = tokenizer.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=True,
    enable_thinking=True,
)
inputs = tokenizer(prompt, return_tensors="pt").to(device)
with torch.inference_mode():
    outputs = model.generate(
        **inputs,
        max_new_tokens=512,
        do_sample=False,
        use_cache=True,
    )
response = tokenizer.decode(
    outputs[0, inputs.input_ids.shape[1]:],
    skip_special_tokens=True,
)
print(response)

The example enables thinking mode, uses greedy decoding, and generates up to 512 new tokens. Set enable_thinking to False to disable thinking, or change device to select a GPU. trust_remote_code=True loads the custom SenseNova Loop model implementation.

Deployment

The following commands are adapted from the Preview deployment examples. Replace the model path with the release checkpoint directory; the release engine versions and configurations will be updated with the model.

LightLLM

Four-GPU deployment example:

docker pull ghcr.io/modeltc/lightllm:main

docker run --rm --gpus all --ipc=host -p 8000:8000 \
  -v /path/to/SenseNova-Loop-9B:/model \
  -e LOADWORKER=4 -e LIGHTLLM_TRITON_AUTOTUNE_LEVEL=1 \
  ghcr.io/modeltc/lightllm:main \
  python -m lightllm.server.api_server \
    --model_dir /model --model_name SenseNova-Loop-9B \
    --tp 4 --host 0.0.0.0 --port 8000

Once the service is running, send an OpenAI-compatible request:

curl http://localhost:8000/v1/chat/completions \
  -H 'Content-Type: application/json' \
  -d '{"model": "SenseNova-Loop-9B", "messages": [{"role": "user", "content": "Hello!"}], "max_tokens": 128}'

SGLang

Use the version from SGLang PR #47281, which includes the SenseNova Loop implementation, and replace the model path:

git clone https://github.com/sgl-project/sglang.git
cd sglang
git fetch origin pull/47281/head
git checkout -b sensenova-loop FETCH_HEAD
python -m pip install -e ./python

sglang serve --model-path /path/to/SenseNova-Loop-9B

vLLM

Use the version from vLLM PR #64832, which includes the SenseNova Loop implementation, and replace the model path:

git clone https://github.com/vllm-project/vllm.git
cd vllm
git fetch origin pull/64832/head
git checkout -b sensenova-loop FETCH_HEAD
python -m pip install -e .

vllm serve /path/to/SenseNova-Loop-9B \
  --served-model-name sensenova_loop \
  --tensor-parallel-size 4 \
  --dtype bfloat16

Inference Acceleration

SenseNova-Loop supports MTP speculative decoding. DSpark and DFlash2 draft model training instructions are available in SenseSpec; available release draft weights and launch settings will be updated with the model.

🏋️ Training

For model fine-tuning, refer to the SenseNova-Loop examples in OpenSenseNova/torchtitan.

Training scripts for DSpark and DFlash2 speculative decoding draft models are available in OpenSenseNova/SenseSpec.

⚠️ Known Limitations

Given its size and development time, SenseNova-Loop-9B has room for improvement in the following areas:

  • Complex instruction following: The model may misunderstand or miss requirements when given many complex tool definitions or multiple constraints.
  • Coding and general agent tasks: Performance can be inconsistent on coding and general agent tasks. The model can correct some errors through additional steps, but its ability to handle complex, long-horizon tasks still needs improvement.
  • Complex web generation: Websites with complex 3D scenes and rich animations may not be completed in a single turn. Some examples require user feedback and two to three rounds of refinement.

We will continue to improve the model and plan to apply the Loop architecture to larger models in the future. We welcome community feedback and reports of issues.

📄 License and Citation

The model project is licensed under the Apache License 2.0.


Downloads last month
-
Safetensors
Model size
9B params
Tensor type
BF16
·
F32
·
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 1 Ask for provider support

Model tree for sensenova/SenseNova-Loop-9B

Adapters
1 model
Finetunes
3 models
Quantizations
14 models

Space using sensenova/SenseNova-Loop-9B 1

Collection including sensenova/SenseNova-Loop-9B