SenseNova Loop 9B Preview — Q4_K_M + F16 vision projector

Unofficial GGUF release. The original model and vision weights are by SenseNova / OpenSenseNova. The text GGUF is requantized from notfakerv's text-only Q8_0 GGUF; the F16 vision projector is converted from the original SenseNova vision weights. This is not an official SenseNova, notfakerv, or llama.cpp release.

SHA256SUMS checks the two GGUF files. See NOTICE and LICENSE for provenance and license terms. The runtime requires all three patches in order on llama.cpp commit ff0d396e3dda0f33e7ce5ad4267fce4f7690cfb1; stock llama.cpp will not load this text model. The third patch only touches conversion/__init__.py and conversion/qwen3vl.py. It reuses the existing Qwen3VL mtmd graph and does not modify the Loop/mHC text graph.

Run

Build the patched llama.cpp with CUDA (on the tested CUDA 13.3 host, set -DCMAKE_CUDA_HOST_COMPILER=/usr/bin/g++-15 -DCMAKE_CUDA_ARCHITECTURES=86):

git clone https://github.com/ggml-org/llama.cpp.git
cd llama.cpp
git checkout ff0d396e3dda0f33e7ce5ad4267fce4f7690cfb1
git apply /path/to/this-repo/runtime/01-clover-pr31306.patch
git apply /path/to/this-repo/runtime/02-sensenova-loop-native.patch
git apply /path/to/this-repo/runtime/03-sensenova-loop-vision.patch
cmake -S . -B build -DGGML_CUDA=ON -DCMAKE_CUDA_HOST_COMPILER=/usr/bin/g++-15 -DCMAKE_CUDA_ARCHITECTURES=86
cmake --build build --target llama-server llama-cli llama-quantize -j 12

Download both GGUFs from this repository, then start:

./build/bin/llama-server \
  -m /path/to/SenseNova-Loop-9B-Preview-Q4_K_M.gguf \
  --mmproj /path/to/mmproj-SenseNova-Loop-F16.gguf \
  -ngl 99 -c 8192 --parallel 1 --port 8080

The built-in web UI accepts images. The OpenAI-compatible /v1/chat/completions API accepts a user message containing text and image_url parts; for a local image, provide a data:image/png;base64,... URL. Text-only usage needs just -m, without --mmproj.

Reproduce the GGUFs

Pinned text source: notfakerv GGUF revision b54ee83643a4e8066290d3062e1a218139c4c007. Requantization from Q8_0 requires --allow-requantize and can reduce quality relative to quantizing from BF16:

hf download notfakerv/SenseNova-Loop-9B-Preview-GGUF SenseNova-Loop-9B-Preview-Q8_0.gguf --revision b54ee83643a4e8066290d3062e1a218139c4c007 --local-dir source-q8
./build/bin/llama-quantize --allow-requantize source-q8/SenseNova-Loop-9B-Preview-Q8_0.gguf SenseNova-Loop-9B-Preview-Q4_K_M.gguf Q4_K_M 12

Pinned original model revision: 27abea779b1e6634b136805b6519371739d551f7. All 333 vision tensors reside in shard 8. Download that shard and the configuration into a new directory without model.safetensors.index.json; the existing converter will index the available shard and discard its text tensors. The entire BF16 checkpoint is not needed and was not run:

hf download sensenova/SenseNova-Loop-9B-Preview config.json preprocessor_config.json processor_config.json tokenizer_config.json generation_config.json README.md model-00008-of-00008.safetensors --revision 27abea779b1e6634b136805b6519371739d551f7 --local-dir vision-source
python convert_hf_to_gguf.py vision-source --mmproj --outtype f16 --outfile mmproj-SenseNova-Loop-F16.gguf

Validation and limits

Tested on an RTX 3060 12 GB at 8192 context (about 9.9 GB VRAM while serving): text-only 13 × 7 → 91; image recognition star → Star; the same color question changed from Red to Blue when only the image was flipped; and counting three circles then adding two gave 5. An image request through /v1/chat/completions returned 200 and the correct color. I also tried a video clip, and the model described it quite accurately. I haven't tested video support more broadly yet.

Image resizing metadata follows the original HF processor (16-pixel patches, 2×2 spatial merge, 65,536 minimum and 16,777,216 maximum pixels). Very large images can exceed an 8192-token context; reduce the image or set an appropriate image-token limit. Numeric comparison against the complete HF BF16 model was not performed. Multi-image, general video behavior, long-context quality, vision quantization, MTP, diffusion acceleration, and speculative decoding are not validated by this release.