SenseTime released SenseNova-Loop-9B on 2026-11-17 under the Apache 2.0 license. It is an open-weight multimodal model that accepts text and images and returns text, with about 9.4B parameters and a context window of up to 512K tokens. A looped architecture runs 24 shared decoder layers twice, giving 56 layer executions from 32 physical layers, and it is built for reasoning, coding, tool use and data analysis. The official 4-bit AWQ weights take about 4.9 GB, and 256K-context inference needs roughly 15 GiB of GPU memory.
A solid 9.41B-parameter dense language model from sensenova. A pragmatic middle-ground choice when you need open weights without a flagship-sized footprint. Newly released, so production-readiness is still being shaken out.
Generated from this model’s benchmarks and ranking signals. Editor reviews refine it over time.
Access model weights, configuration files, and documentation.
No benchmark data available for this model yet.
Benchmark data sourced from: Hugging Face, Arena.ai, Artificial Analysis.
The top devices for this model at 4-bit, ranked by fit and speed.
| Device | Grade | Speed | VRAM |
|---|---|---|---|
| NVIDIA H100 SXM5 80GBNVIDIA | SS | 60.6 tok/s | 44.5 GB |
| Google Cloud TPU v5pGoogle | SS | 50.0 tok/s | 44.5 GB |
| Intel Gaudi 2 AI AcceleratorIntel | SS | 44.3 tok/s | 44.5 GB |
| NVIDIA A100 SXM4 80GBNVIDIA | SS | 36.9 tok/s | 44.5 GB |
| Intel Gaudi 3 AI Accelerator |

Explore the Provider
Aggregate stats, leaderboard, release timeline, and benchmark coverage across every sensenova model we track.
| SS |
| 66.9 tok/s |
| 44.5 GB |
See how different quantization levels affect VRAM requirements and quality for this model.
| Format | VRAM Required | Quality | |
|---|---|---|---|
| Q2_K | 42.5 GB | Low | |
| Q4_K_MRecommended | 44.5 GB | Good | |
| Q5_K_M | 45.4 GB | Very Good | |
| Q6_K | 46.6 GB | Excellent | |
| Q8_0 | 48.9 GB | Near Perfect | |
| FP16 | 57.9 GB | Full |
SenseNova-Loop-9B is an open-weight, multimodal vision-language model developed by SenseTime, published under the permissive Apache 2.0 license. Operating at 9.41B physical parameters, the model is engineered to punch well above its weight class by deploying a looped execution mechanism. Rather than bloating disk and memory footprints with 14B to 27B raw weights, SenseNova-Loop-9B stores 32 physical decoder layers but cycles through its middle layers twice during each forward pass. The result is a parameter-efficient network that delivers the effective depth of a 14.6B parameter model while maintaining the static memory profile of a sub-10B architecture.
For developers seeking to run SenseNova-Loop-9B locally, this design introduces a distinct engineering tradeoff: lower VRAM overhead at the cost of additional compute cycles per generated token. Equipped with an expansive 512,000-token context window, native vision input, and dedicated optimizations for tool calling and reasoning, the model directly challenges larger open-weight models such as Qwen2.5-Coder-14B and Qwen3.8-27B on complex agentic and analytical workloads. It bridges the gap between lightweight edge deployment and deep multi-step inference.
As a local AI model 9.41B parameters 2026 release, SenseNova-Loop-9B delivers high utility for developers building self-hosted autonomous agents, code-generation pipelines, and multimodal document processors. Its ability to inspect high-resolution images, digest massive code repositories, and produce structured tool calls locally makes it a standout choice for air-gapped environments and constrained hardware setups.
The core innovation of SenseNova-Loop-9B lies in its looped transformer decoder. Traditional dense models map every layer execution to unique parameter storage on disk and in VRAM. In contrast, SenseNova-Loop-9B utilizes 32 physical layers to generate 56 layer executions per forward pass:
Because the model runs 56 layer operations over 32 stored layers, its computational complexity per token aligns with a 14.6B model, even though only 9.41B parameters reside in memory. This means your GPU memory bandwidth and compute cores are worked harder per token than in a standard 9B model, but the static VRAM requirement remains low.
SenseNova-Loop-9B is natively multimodal, accepting interleaved text and image inputs while outputting text. SenseTime explicitly targeted reasoning, coding, and multi-step tool execution during post-training:
The model is tuned for autonomous agent environments requiring reliable JSON formatting, API calls, and self-correction loops. On agentic benchmarks like NL2Repo (scoring 41.36) and AIDABench (scoring 65.31), SenseNova-Loop-9B competes directly with larger models like Qwen3.8-27B. It can manage multi-turn state machines, inspect environmental output, and generate valid function calls without drift.
Unlike language-only models that require a separate visual encoder pipeline, SenseNova-Loop-9B processes image tokens natively. It handles dense visual artifacts including complex circuit diagrams, mechanical beam schematics, handwritten mathematical equations, and scanned financial receipts. This makes it suitable for automated invoice extraction, OCR verification, and local visual QA pipelines.
In software development tasks, the looped architecture aids long-horizon dependency tracking. The model can parse multi-file repositories within its 512K context, isolate bugs across disparate modules, generate complex SQL queries from schema diagrams, and write complete test suites. Its recursive layer passes provide extra computational depth, helping eliminate logic errors in algorithmic code generation.
Deploying the model locally requires understanding the relationship between weight storage, runtime compute, and the KV cache needed for long-context execution.
The weights themselves are remarkably compact, but active memory consumption depends heavily on how much of the 512K context window you utilize:
For the majority of developers, 4-bit AWQ or GGUF Q4_K_M represents the optimal sweet spot. Because the looped architecture cycles through layers 4 to 27 twice, quantization errors in those specific shared weights compound across passes. High-quality quantization schemes like AWQ preserve activation outliers in the shared blocks far better than naive round-to-nearest methods, maintaining near-FP16 reasoning performance.
Because every token requires 56 layer evaluations rather than 32, SenseNova-Loop-9B tokens per second metrics will look lower than a traditional 9B model:
num_ctx) to prevent uncontrolled VRAM allocation when loading massive documents.To evaluate whether SenseNova-Loop-9B fits your stack, compare it against its closest competitors in the 8B to 14B open-weights category:
When hardware memory is your primary constraint, but your agentic workflow demands deep multi-step reasoning, tool execution, and visual inputs, SenseNova-Loop-9B provides an exceptionally capable architecture for local deployment.
Energy cost on Corsair AI Workstation 300 (Ryzen AI Max 385) (~4.6 tok/s, Q4_K_M) vs flagship API pricing.
| Source | Cost per 1M tokens |
|---|---|
Local (energy only)SenseNova-Loop-9B on Corsair AI Workstation 300 (Ryzen AI Max 385) · ~4.6 tok/s · 150W | $1.08 |
GPT-6.2 SolOpenAI · in $2.00 · out $10.00 | $4.40 |
Claude Fable 5.2Anthropic · in $8.00 · out $40.00 | $17.60 |
Gemini 4 NeonGoogle · in $3.00 · out $15.00 | $6.60 |
Grok 4.5xAI · in $2.00 · out $6.00 | $3.20 |
API prices blended at 70% input / 30% output.
Hardware amortisation not included. Run the full ROI calculator for payback math.
Cheapest current cloud rentals with at least 45 GB VRAM, refreshed hourly.
Advertising disclosure: we earn commissions when you shop through the links below.
| Option | Cost / GPU-hour |
|---|---|
NVIDIA A100 80GB PCIeVast.ai · Spot · 80 GB VRAM | $0.16 |
NVIDIA A40Vast.ai · Spot · 48 GB VRAM | $0.27 |
NVIDIA A40Vast.ai · On-Demand · 48 GB VRAM | $0.27 |
NVIDIA RTX A6000RunPod · Community · 48 GB VRAM | $0.33 |
NVIDIA L40SVast.ai · Spot · 48 GB VRAM | $0.40 |
Per-GPU rate across RunPod, the Vast.ai marketplace, DigitalOcean, and Vultr.
Spot tier is interruptible. Plan for restarts when comparing against on-demand prices.