Can I Run SenseNova Loop 9B Locally? (2026): Patched Only

Bottom Line

• Yes, on an 8GB card, but only with a patched llama.cpp. The Q4_K_M file is 5.63 GB. Q8_0 is 9.54 GB and fits a 16GB card with room for context.
• Stock runtimes do not load it. On 4 November 2026 the GGUF card says upstream llama.cpp, LM Studio and Ollama “do not support this architecture yet.” Issue #31571 asks for it; no pull request exists.
• Expect 14B-class speed from 9B of memory. The model runs its middle 24 layers twice per token. That is the opposite trade from a mixture-of-experts model.
• Text-only for now. The GGUFs ship without the vision tower (mmproj). The multimodal path needs Transformers, LightLLM, SGLang or vLLM.
• Action: if you want to test a looped model today, build the patched llama.cpp and run Q8_0 on 16GB. If you need Ollama or LM Studio, wait for upstream support.

Ready to buy? See the tested hardware list with current prices.

What SenseNova Loop 9B Is

	SenseNova-Loop-9B-Preview
Maker	SenseTime (SenseNova)
On Hugging Face	30 October 2026
License	Apache-2.0
Parameters	about 9B stored, about 14.6B at inference
Distinct layers	32
Looped block	layers 4 to 27, run twice: 56 layer executions per forward pass
Attention	three linear-attention layers, then one full-attention layer, per group of four
Full-attention heads	16 query heads, 4 KV heads
Hidden size	4,096 (feed-forward 12,288)
Vocabulary	248,320
Residual streams	two mHC streams, 20 Sinkhorn iterations
Training context	up to 262,144 tokens; documented recipes go to 256K
Export precision	BF16 backbone, FP32 mHC routing parameters
Input	text and images (official weights); text only (GGUF)

Why it trends. SenseTime open-sourced the model on 30 October 2026. AICoder and insideAI covered it the same day, aicybr published a local-inference guide on 1 November, and MindStudio ran a hands-on review on 2 November. Seven community GGUF repositories appeared on Hugging Face by 4 November; the AtomicChat repository alone had 9.63k downloads when we read it.

Every GGUF Size

Sizes are what the Hugging Face file tree shows, in GB, read 4 November 2026. The AtomicChat card’s own table lists the same files smaller (Q4_K_M 5.2 GB, Q8_0 8.9 GB, BF16 16.7 GB) because it counts in GiB.

Repo	Quant	Size (GB)
guideboardlabs	Q3_K_M	4.64
guideboardlabs	Q4_0	5.33
AtomicChat / guideboardlabs	Q4_K_M	5.63
AtomicChat	Q5_K_M	6.47
AtomicChat	Q6_K	7.37
AtomicChat	Q8_0	9.54
AtomicChat	BF16	17.9

The full BF16 safetensors repository is about 38.2 GB, per the aicybr guide. That checkpoint is for the Transformers, LightLLM, SGLang and vLLM paths. The GGUF files are the only route for a single consumer card.

Why It Runs Like a 14B Model

SenseNova Loop stores 32 distinct decoder layers. At inference it runs layers 4 to 27 twice, so each token passes through 56 layer executions. That is where the “about 14.6B inference parameters” figure comes from: 9B of weights, 14.6B worth of compute.

This is the opposite trade from a mixture-of-experts model. A MoE such as Clover 35B-A3B stores 35B parameters and touches about 3B per token: big memory, fast decode. SenseNova Loop stores 9B and computes 14.6B: small memory, slower decode. On a desktop GPU, decode speed scales with the bytes read per token, so expect speed closer to a dense 14B model than to a dense 9B model.

The KV cache is harder to size. The model has 8 full-attention layers among its 32 distinct layers, and the looped block contains full-attention layers that execute twice. The card does not say whether the two passes share one cache or keep two. We do not publish KV arithmetic for this model until the runtime documents it. Reserve 1 to 2 GB above the file size for a 32K context and measure.

What Fits Your Machine

Machine	File to run	Notes
8GB card	Q4_K_M (5.63)	The GGUF card says Q4_K_M “runs on ~8 GB-class GPUs.” Short context.
12GB card	Q6_K (7.37)	Room for 32K context.
16GB card	Q8_0 (9.54)	The comfortable pick. Near-lossless, the card says it “fits comfortably in 16 GB.”
24GB card	BF16 (17.9)	Full precision in GGUF form.
Mac, 16GB	Q4_K_M or Q5_K_M	Leave room for macOS. No Mac number has been published.
Mac, 32GB+	Q8_0 or BF16	Untested by anyone we can cite.

For 16GB, a new RTX 5060 Ti 16GB holds Q8_0 with 6 GB spare; check current listings for its price. For 24GB, a used RTX 3090 holds BF16 and cost about $1,500 as of November 2026 (lowest ask on GPU Dojo, 3 November).

A 9B file does not need a 24GB card. Buy the 24GB card for the other models in the 24GB tier, not for this one.

How Fast It Runs

One local number exists, and it is not ours.

Hardware	Setup	Generation	Source
NVIDIA A6000	long multimodal task, flash attention, VRAM about 18.5 GB	about 2 tok/s	MindStudio hands-on, 2 November 2026 (community-reported)

Read it with three caveats:

1. The review does not name the runtime. The model ships paths for Transformers, LightLLM, SGLang and vLLM; the review says some of those are open pull requests. A Transformers path at BF16 is not a llama.cpp number.
2. It is a preview release. The review says the model has no finalized technical report and no optimized inference speed yet.
3. No llama.cpp, LM Studio or Ollama number exists. The GGUF card publishes none. We do not estimate one.

Run It

Status on 4 November 2026:

• Upstream llama.cpp does not load it. The AtomicChat card: “New architecture — needs a patched llama.cpp.” The files “were produced and verified with a llama.cpp build that adds SenseNova-Loop support, on top of ggml-org/llama.cpp commit 5047748f11 (upstream support PR pending).” Our search of llama.cpp pull requests found no open PR; issue #31571, opened 3 November, is the request.
• LM Studio and Ollama do not load it. The card says so directly. Ignore the auto-generated ollama run and vllm serve snippets in the Hugging Face sidebar.
• Text only. “These are text-only GGUFs for now; the vision tower (mmproj) is a follow-up.”
• The reference path is Python. Transformers 5.18.0 with PyTorch 2.11.0 and CUDA 13.0, with trust_remote_code=True. LightLLM’s documented example uses four-way tensor parallelism. SGLang support is PR #47281; vLLM support is PR #64832, run with --tensor-parallel-size 4 --dtype bfloat16.

If you build the patched tree the GGUF author used, the server command is the normal one. This example is for an NVIDIA card; on a Mac, drop the CUDA flag.

# needs a llama.cpp build with sensenova_loop support (not upstream on 2026-11-04)
cmake -B build -DGGML_CUDA=ON && cmake --build build -j --config Release
./build/bin/llama-server -m SenseNova-Loop-9B-Preview-Q8_0.gguf \
  -ngl 99 -c 32768 -fa on --jinja

Check your flags against llama.cpp flags explained first. The -lm change broke most copied commands in 2026.

Should You Try It?

Try it if you want to test the looped-transformer idea on a card you already own. The memory cost is a 9B model. The benchmarks on the card are a single chart image with no numbers in the text, so we cannot compare it with Qwen or Gemma here.

Wait if you need a stock runtime, vision input, or a measured speed. Every one of those is missing on 4 November 2026.

Sources

• Model card: sensenova/SenseNova-Loop-9B-Preview (read 2026-11-04)
• GitHub: OpenSenseNova/SenseNova-Loop (read 2026-11-04)
• GGUF sizes and support status: AtomicChat/SenseNova-Loop-9B-Preview-GGUF, guideboardlabs/SenseNova-Loop-9B-Preview-GGUF (read 2026-11-04)
• llama.cpp issue #31571, opened 2026-11-03
• Hands-on speed: MindStudio review, 2026-11-02
• Checkpoint size and VRAM planning: aicybr guide, 2026-11-01
• Press: AICoder, 2026-10-30
• RTX 3090 price: GPU Dojo lowest ask, read 2026-11-03

See Also

• Can I run Clover 35B-A3B locally? — the MoE that makes the opposite memory-for-speed trade
• Best local LLM for 16GB VRAM — what else the Q8_0 tier holds
• Best local LLM for 24GB VRAM — the tier that holds the BF16 file
• llama.cpp flags explained — the 2026 flag changes before you paste a command
• Can I run Nanbeige4.3 33B locally? — a dense 33B at the same preview stage

Before you order parts, check the tested hardware list for current prices by tier.