SenseNova-Loop-9B-AWQ 512K Context Multimodal Model: GGUF Builds

November 19, 2026

SenseNova-Loop-9B-AWQ 512K Context Multimodal Model: GGUF Builds

At a Glance

Item Value
Repository sensenova/SenseNova-Loop-9B-AWQ
Family guide SenseNova-Loop guide (2 articles)
Published 2026-11-18
License apache-2.0
Formats safetensors
Source type Primary source (the publisher itself)

Values determined by this site’s code when the information was collected. Dates are JST.

What we checked ourselves

  • It needs about 21% fewer tokens than the Qwen3 tokenizer for the same Japanese text, so about 1.26× as much Japanese fits in the same context length, and generation is faster per character.

Details and conditions are in “Our Own Measurements” below.

Overview

sensenova/SenseNova-Loop-9B-AWQ, the official 4-bit AWQ quantized version of sensenova/SenseNova-Loop-9B, has been released. This model is a multimodal model that accepts text and image inputs and generates text.

By combining a looped architecture that reuses layer weights with manifold-constrained hyper-connections (mHC), it expands the depth of computation while maintaining a compact parameter scale. It is designed for applications such as complex reasoning, coding, tool use, agent tasks, and data analysis. sensenova/SenseNova-Loop-9B-AWQ

Specifications

  • Parameters: Approximately 9B (approx. 9B actual parameters, equivalent to approx. 14.6B in inference computational scale)
  • Architecture: Looped multimodal decoder (32 distinct decoder layers, 56 layer executions per forward pass, two-stream mHC, hybrid attention)
  • Context length: 512K tokens

Performance

The following are the configuration tables and performance information of the original model (BF16 version) as stated in the model card. Minor performance variations may occur in the quantized version due to quantization.

Model Configuration Specification Table

The configuration and mHC settings described in the model card are as follows:

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
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

Benchmark Evaluation and Comparison with Other Models

According to the published values on the model card, the comparison scores between this model (Release version), the previous version (Preview version), and other models are as follows:

  • AutomationBench: SenseNova-Loop-9B Release (36.67) / Preview (14.33) / Qwen3.8-27B (36.67)
  • Terminal-Bench 2.1: SenseNova-Loop-9B Release (59.55) / Preview (40.45)
  • GAIA: SenseNova-Loop-9B Release (69.90) / Preview (37.86)
  • AIDABench: SenseNova-Loop-9B Release (65.31) / Qwen3.8-27B (65.01)
  • NL2Repo: SenseNova-Loop-9B Release (41.36) / Qwen3.8-27B (44.60)
  • BIRD: SenseNova-Loop-9B Release (72.34) / Preview (71.38)

According to the publisher’s measurement results, scores have improved across all evaluation items compared to the previous Preview version. In particular, GAIA (accuracy in solving real-world research tasks using tools such as search and file operations) increased significantly from 37.86 to 69.90. Furthermore, Terminal-Bench 2.1 (accuracy in actually executing commands in a terminal and seeing tasks through to completion) also improved from 40.45 to 59.55, indicating enhanced execution capabilities in agent tasks.

In comparison with the larger-scale Qwen3.8-27B, it shows equivalent or better performance on AIDABench (65.31 vs 65.01) and AutomationBench (36.67 vs 36.67). On the other hand, its score on NL2Repo is 41.36, falling short of Qwen3.8-27B’s 44.60. It should be noted that the range where strengths are demonstrated varies depending on the nature of the task.

Strengths and Use Cases

This model is a multimodal model that accepts both text and image inputs and generates text. It inherits the features of the original model, sensenova/SenseNova-Loop-9B, supporting an extra-long context length of up to 512K tokens. This makes it suitable for understanding very long contexts and executing multi-step tasks.

According to the official model card, the model is primarily designed for the following domains and applications:

  • Agent tasks and data analysis: Excels at practical problem processing involving automated command execution on terminals, tool manipulation, and data analysis.
  • Coding and tool use: Supports tasks incorporating code writing, programming problem solving, and external tool calls.
  • Multimodal content and web design: In addition to processing based on input images, it supports generation involving visual elements such as 3D scene descriptions, SVG code output, and poster or interactive website design.

On the other hand, the model card lists the following points as notes on use and current limitations:

  • Complex instruction following: When numerous tool definitions or multiple constraints are given simultaneously, requirements may not be correctly recognized or may be overlooked.
  • Consistency in long-term agent tasks: Behavior may vary in coding and agent operations, and there is room for improvement in the ability to complete complex tasks with many steps and long durations.
  • Complex web generation: For web content generation involving advanced 3D expressions and animations, completion cannot be achieved with a single instruction, requiring multiple rounds of dialogue and corrections with the user.

How to Get It

sensenova/SenseNova-Loop-9B-AWQ is distributed in the 4-bit AWQ quantized Safetensors format. Since it is not a gated model that requires agreeing to terms of use or submitting a request, it can be downloaded and used directly from Hugging Face.

Using this model requires an inference runtime that supports this AWQ layout. Deployment is supported on the inference engine LightLLM, and an API server can be launched using a single GPU by specifying the local path where the repository was downloaded and running the following command:

python -m lightllm.server.api_server \
  --model_dir /path/to/SenseNova-Loop-9B-AWQ \
  --model_name sensenova_loop \
  --tp 1 \
  --port 8000 --running_max_req_size 4

In addition to the main model weights located in the root of the repository, draft model weights usable for speculative decoding features are located in the mtp/ directory as an option.

Can You Run It Locally?

Runs in Ollama, LM Studio and llama.cpp via a converted build.

The publisher ships safetensors, but AtomicChat/SenseNova-Loop-9B-GGUF provides a GGUF build you can use.

License — apache-2.0 (Commercial use allowed): Permits commercial use, modification and redistribution. Redistribution requires including the license and stating changes; includes a patent grant.

Compiled by this site’s code from the published formats, converted builds we have found, and each engine’s own model registry. “Not found" means we have not seen such a build, not that none exists. License summaries are not legal advice — check the publisher’s original terms before relying on them.

Our Own Measurements

Values we measured ourselves on our server (no GPU) by actually reading and running this model’s files — not figures copied from the model card. How we measure, and the results for every model: Our Measurements.

What We Found

SenseNova-Loop-9B features high token efficiency for Japanese, requiring fewer tokens to represent 1,000 characters compared to Qwen3. The ratio of tokens for Japanese relative to English is nearly one-to-one, indicating efficient context handling across languages. However, it cannot be run directly on local CPU environments at this time because the model architecture is not yet supported by llama.cpp. Future updates to llama.cpp are expected to enable the use of this model, allowing users to leverage its high token efficiency.

Japanese Token Efficiency

Tokenizer Tokens per 1,000 Japanese characters Ratio to the same text in English
This model 546 0.99×
Qwen3 688 1.26×
Llama 3.2 744 1.36×
Gemma 3 564 1.03×
gpt-oss 795 1.45×
LLM-jp-3 497 0.85×

It needs about 21% fewer tokens than the Qwen3 tokenizer for the same Japanese text, so about 1.26× as much Japanese fits in the same context length, and generation is faster per character.

Counted with the tokenizer.json of sensenova/SenseNova-Loop-9B-AWQ on a fixed text we wrote ourselves (876 Japanese characters across news, conversation, technical docs, a formal email, travel writing and a recipe) and its English translation. Fewer tokens mean more Japanese fits in the context window.

Inside the GGUF File

File: SenseNova-Loop-9B-AD-Q4_K_M.gguf (5.55GB, Q4_K_M). We read only the header (metadata) of the file, not the weights.

Item Value
Architecture (as named in the GGUF) sensenova_loop
Maximum trained context length 524,288 tokens
Layers 57
Vocabulary size 248,320
Chat template Included (mentions tool calls, has a thinking switch)
imatrix Used (calib_train.txt, 19,368 chunks)
Weight types (share of parameters) Q5_K 39.8% / Q4_K 35.7% / Q6_K 12.5% / IQ4_XS 11.1% / other 0.8%
Average bits per weight 5.17 bits
Embedding / output layer type IQ4_XS / Q6_K

The quant name in the file name describes the file as a whole; in practice layers mix several types. The average bits per weight is the measured file data divided by the number of weights.

Running It on a CPU Only

llama.cpp b12418 could not load this file: it does not know the architecture sensenova_loop yet. We will measure it again once support lands.

Conditions: Neoverse-N1, 3 threads, no GPU, llama.cpp b12418, Q4_K_M (5.55GB) from AtomicChat/SenseNova-Loop-9B-GGUF, measured 2026-11-19 (JST).

Quantized and Converted Variants

→ Scroll horizontally to see all columns

Added Publisher Format Repository Smallest VRAM tier (build, est. memory)
2026-11-18 AtomicChat GGUF (imatrix) AtomicChat/SenseNova-Loop-9B-GGUF Q5_K_M 7.7GB (fits in 8GB VRAM)

File sizes of each build:

  • Available builds in AtomicChat/SenseNova-Loop-9B-GGUF: IQ3_S 4.5GB / IQ4_XS 5.4GB / Q4_K_M 5.6GB / Q5_K_M 6.4GB / Q6_K 8.2GB

This section is appended automatically by Local Model Watch when a converted build of this model appears after publication. Memory figures are estimated from the size of the distributed files. What the quantization names mean: glossary.

Related Articles

What to Read Next

Sources

Update History

  • 2026-11-19: Added converted builds to “Quantized and Converted Variants”: AtomicChat/SenseNova-Loop-9B-GGUF
  • 2026-11-19: Added our own measurements: Japanese token efficiency, what is inside the GGUF file.
  • 2026-11-19: Changed the title to show what the article covers (VRAM requirements, file list, etc.).