# SenseNova-Loop-9B-Preview (slim)

> SenseNova-Loop-9B-Preview is SenseTime's open-weight multimodal model of about 9B parameters. It runs 24 of its 32 decoder layers twice, with text and images in, tool calling, a thinking switch and a 256K context. Apache 2.0. No released runtime loads it.

- Full: https://www.anchorterminal.com/tools/sensenova-loop-9b-preview.md (~10,250 tokens) · this version ~1,730 tokens · JSON https://www.anchorterminal.com/tools/sensenova-loop-9b-preview.json · canonical https://www.anchorterminal.com/tools/sensenova-loop-9b-preview
- Index: https://www.anchorterminal.com/llms.txt · API: https://www.anchorterminal.com/api/v1/index.json · MCP: https://www.anchorterminal.com/mcp · Updated: 2026-11-10 · pricing https://www.anchorterminal.com/tools/sensenova-loop-9b-preview#pricing · reviews https://www.anchorterminal.com/tools/sensenova-loop-9b-preview#reviews
- Licence: CC BY 4.0, credit Anchor Terminal and link this page (https://www.anchorterminal.com/terms/#data)

**D · 46.8/100 · rank #155 of 160 open-weight models · #110 in Open-weight models · not agent-ready · confidence low**


Assessment: An unusual architecture under Apache 2.0, with a 256K training context, a thinking switch and tool calling in the template. It's a preview in practice too: no released runtime runs it, loading it runs the repository's code, the files are twice the size the card implies, and nobody from SenseNova has answered a question yet.

## Facts

- Kind: Open-weight model · vendor: SenseTime (SenseNova) · category: Open-weight models · legal entity: SenseTime, listed on the Stock Exchange of Hong Kong (00020) per its investor page. SenseNova is its foundation model series, per its site · provenance 51/100
- Packages: huggingface `sensenova/SenseNova-Loop-9B-Preview`
- Auth: None · pricing: Free · x402: no · licence: Apache-2.0
- Benchmark probes: not run yet, so Performance is pending
- Surface graded: The open weights at huggingface.co/sensenova/SenseNova-Loop-9B-Preview, read from the model card, the files and their metadata, the config, the chat template, the commit history, the README and LICENSE of the GitHub repository and the runtimes' own code at their latest releases on 10 November 2026. Who SenseNova is comes from SenseTime's home, about and investor pages. The repository's model code, the training guides and the two pull requests weren't read. Nothing was downloaded or run
- Model: 9,656,250,480 parameters in the files. The card says about 9B in total and about 14.6B at inference, since layers 4 to 27 of its 32 decoder layers run twice with shared weights. SenseNovaLoopForConditionalGeneration, with linear and full attention interleaved and two residual streams in the looped section. The config's text and vision parts carry Qwen 3.5's type names
- Files: Eight safetensors shards (37.65 GB) and mtp.safetensors (0.49 GB) for multi-token prediction, chat_template.jinja, config, processor and tokenizer files, and modeling_sensenova_loop.py, the model's own code. No pickle files. Hugging Face lists a SHA-256 for the large files
- Licence: Apache 2.0 in the card's metadata, and the card links the LICENSE in the GitHub repository. The model repository has no LICENSE file of its own
- Runtimes: Transformers 5.18.0 with trust_remote_code, per the card. vLLM and SGLang only from open pull requests (vLLM 64832, SGLang 47281), and LightLLM, a server from ModelTC, on four GPUs. Released vLLM 0.33.0, SGLang 0.5.23, Transformers 5.22.2 and llama.cpp b12242 have no sensenova_loop code
- Context: Up to 262,144 tokens in training, per the card's specification table
- Hardware: No memory figure on the card. Its Transformers example uses one GPU, and its vLLM and LightLLM examples four
- Training data: One sentence: the documented recipes combine long sequences, multimodal inputs and supervised task trajectories. No sources, sizes or cut-off, and the technical report is marked TBD
- Languages: English and Chinese, per the card's metadata
- Tool calling: The chat template carries a tools block and writes calls as tool_call blocks with function and parameter tags. The card says the training recipes include tool trajectories and gives no tool-calling example or parser
- Reasoning: A thinking mode set with enable_thinking in the chat template call, off in the card's example. The template wraps reasoning in think tags
- Recommended settings: The card's example uses greedy decoding. generation_config.json sets temperature 1.0 and top_p 0.95

- Scores: Capability pending, Reach 40, Risk 50, Agent interface 60, Documentation and upkeep 44 · total over the 7 assessed categories
- Why:  · Reach, Memory 15, the only build is the safetensors set, 38.14 GB with its MTP head, since the Hub counts 9.41B of its parameters in F32. Runtimes… · Risk, File safety 15, safetensors only, no .bin, .pt or .pth files (25). · Agent interface, Chat template 20, chat_template.jinja in the repository (7,756 characters). · Documentation and upkeep, Model card 14, Intended use, the architecture, the parameter counts, the training context and the languages (in the metadata) are on the car…
- Sources: 19, open questions: 6, both in the full twin
- Capabilities: model.open-weights
- JSON: https://www.anchorterminal.com/api/v1/tools/sensenova-loop-9b-preview.json
- Its vendor verifies the listing with `verify_listing` or `POST https://www.anchorterminal.com/api/v1/verify` (the full twin says how); no effect on the grade.
- Support: https://github.com/OpenSenseNova/SenseNova-Loop; more in the full twin

## Before you call it

1. Read modeling_sensenova_loop.py before you load it. trust_remote_code runs it
2. Pin transformers 5.18.0, the version the card checked
3. Load with dtype=torch.bfloat16 as the card does, or the 38 GB of F32 weights stay at full size
4. Pass enable_thinking=False to apply_chat_template for a direct answer
5. Tool calls come back as tool_call blocks with function and parameter tags. Parse them yourself, since no runtime ships a parser
6. Wait for the vLLM or SGLang pull request to land in a release before putting it behind an API

## Connect

```bash
pip install "transformers==5.18.0" accelerate safetensors
# then AutoModelForCausalLM.from_pretrained("sensenova/SenseNova-Loop-9B-Preview", trust_remote_code=True, dtype=torch.bfloat16)
# trust_remote_code runs modeling_sensenova_loop.py from the repository. Read it first
```

```bash
# No released server runs it. The card's vLLM and SGLang commands build from open pull requests,
# and its LightLLM example serves an OpenAI-compatible API at http://localhost:8000/v1 on four GPUs
```

## Similar tools

| Tool | Grade | Score | Shared capabilities | Slim |
| --- | --- | --- | --- | --- |
| Qwen3-Embedding-8B | BB | 75.4 | model.open-weights | https://www.anchorterminal.com/tools/qwen3-embedding-8b.min.md |
| Gemma 4 31B IT | BB | 74.1 | model.open-weights | https://www.anchorterminal.com/tools/gemma-4-31b-it.min.md |
| Qwen3-Embedding-4B | BB | 73.6 | model.open-weights | https://www.anchorterminal.com/tools/qwen3-embedding-4b.min.md |
| Gemma 4 12B IT | BB | 72.3 | model.open-weights | https://www.anchorterminal.com/tools/gemma-4-12b-it.min.md |
| Gemma 4 26B A4B IT | BB | 72.2 | model.open-weights | https://www.anchorterminal.com/tools/gemma-4-26b-a4b-it.min.md |


## Panel reviews (0, desk reviews from public material, no calls made)