SenseNova-Loop-9B-Preview: Specs, License & How to Run

SenseNova-Loop-9B-Preview is SenseTime’s ungated looped multimodal checkpoint (text and images in, text out). Official hosts we opened on 2026-11-03: the card sensenova/SenseNova-Loop-9B-Preview (created 2026-10-28T15:34:40Z; last modified 2026-11-02T15:43:56Z; 184 likes / 1,536 downloads), config.json, and the companion tree github.com/OpenSenseNova/SenseNova-Loop (86 stars / 7 forks; created 2026-10-29T03:12:37Z; last push 2026-11-02T15:43:48Z; Apache-2.0 LICENSE). API gated was false. Hugging Face pipeline_tag is image-text-to-text; model_type is sensenova_loop. Card news line: released 2026-10-30.

This page hangs off ChatGPT alternatives and closed LLM APIs. That mapping is editorial. We have no sourced quality comparison against ChatGPT. Same-family video card already on this site — not replaced by this page: SenseNova 6.8 Code2Video. Identification vs that 27B code-video checkpoint: SenseNova-Loop-9B-Preview vs SenseNova 6.8 Code2Video.

The card says the model is developed for long-context and agentic workflows. A contiguous block of decoder layers is executed twice with shared weights, while learned mappings manage two residual streams (mHC) in the looped section. Those architecture sentences stay upstream.

What we verified (2026-11-03)

Field	Value	Source
Host	huggingface.co/sensenova/SenseNova-Loop-9B-Preview (not gated)	HF API
Created / last modified	2026-10-28T15:34:40Z / 2026-11-02T15:43:56Z	same
Likes / downloads	184 / 1,536	same
License tag	apache-2.0	HF cardData
LICENSE on the weight repo	404	HF raw fetch
GitHub LICENSE / API	Apache-2.0 file / Apache-2.0	GitHub
Pipeline tag	image-text-to-text	HF API
Architecture	sensenova_loop / SenseNovaLoopForConditionalGeneration	card + config
Params (card)	~9B total; ~14.6B inference parameters	card table
safetensors.parameters	9,656,250,480	HF API
usedStorage	38,204,212,983 bytes (~35.6 GiB)	same
Distinct layers / executions	32 / 56	card
Looped block	layers [4, 28) / 2 passes	card + loop_settings
mHC	2 streams; Sinkhorn 20	card + config mhc
Context	262,144 tokens	card + text_config.max_position_embeddings
Vocab / hidden	248,320 / 4,096	config
GitHub stars / forks	86 / 7	GitHub API
Card news date	2026-10-30	README

text_config.model_type is qwen3_5_text and vision_config.model_type is qwen3_5_vision. We copy those config labels. This is not the Qwen3.8-27B card and not a Qwen3.8-27B fine-tune statement on the Loop README we opened.

The card’s “Key Benchmarks” figure stays on assets/loop9b-benchmark-results.png. We do not OCR scores out of that PNG.

How to run (from the official card, not tested)

Checked Transformers path on the card: Transformers 5.18.0, PyTorch 2.11.0 with CUDA 13.0, trust_remote_code=True. Copied install and generate names only:

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

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "sensenova/SenseNova-Loop-9B-Preview"
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={"": "cuda:0"}
).eval()

trust_remote_code=True loads modeling_sensenova_loop.py. The example uses greedy decoding and enable_thinking=False unless you pass True.

LightLLM on the card uses --tp 4. SGLang is named via PR #47281. vLLM is named via PR #64832 with --tensor-parallel-size 4 --dtype bfloat16. Those are card-named branches, not a claim that the PRs have merged. MTP is listed as available; Dspark and Dflash2 are “training in progress.” forkalt has not run those commands.

Fine-tune docs on the card point at OpenSenseNova/torchtitan and OpenSenseNova/SenseSpec. Those are companion trees, not this weight repo.

License

Hugging Face license: apache-2.0. We did not find a LICENSE blob on the weight repo on 2026-11-03. OpenSenseNova/SenseNova-Loop LICENSE is Apache License 2.0; GitHub’s license API field is Apache-2.0. The card README points at that file. Not legal advice.

What we did not invent

• Numeric BIRD-Dev / WorkBuddy / Terminal Bench rows stay on the README figure. Vendor.
• Peak VRAM, tokens/s, and “best model performance” stay vendor.
• We do not rank Loop-9B against SenseNova 6.8 Code2Video, MiniCPM-V 4.7, Qwen3.8-27B, or ChatGPT.

Related on this site: SenseNova 6.8 Code2Video — SenseNova-Loop-9B-Preview vs SenseNova 6.8 Code2Video — Qwen3.8-27B (27B native VLM; different card) — MiniCPM-V 4.7 — Mage-VL-Cowork (computer-use VLM, different job) — vLLM. Catalog: home.