SenseNova-Loop-9B-Preview — multimodal GGUFs

CPU-quantized GGUF tiers of sensenova/SenseNova-Loop-9B-Preview. Each main GGUF includes the text model and its MTP draft block. For image input, pair any tier with the shared BF16 multimodal projector (mmproj/ below).

Files

Shared image projector:

• mmproj/SenseNova-Loop-9B-Preview-MMProj-BF16.gguf (0.86 GiB)

The BPW values are allocator targets, not guarantees of achieved BPW or exact file size. The same BF16 mmproj works with all eight main-model tiers.

Modality and runtime

The main GGUF retains the SenseNova loop execution plan, mHC routing tensors, and the MTP head at block 56. The vision tower and image projection are in the separate BF16 mmproj file. Image input requires a llama.cpp build with SenseNova-Loop and Qwen3VL projector support; MTP speculative decoding also requires runtime support for this architecture.

Example image prompt:

llama-mtmd-cli \
  -m quant/SenseNova-Loop-9B-Preview-GSQ-RCO-IQ3_S.gguf \
  --mmproj mmproj/SenseNova-Loop-9B-Preview-MMProj-BF16.gguf \
  --image picture.jpg \
  --prompt "Describe this image." \
  -ngl 0

Example MTP-enabled text generation:

llama-cli \
  -m quant/SenseNova-Loop-9B-Preview-GSQ-RCO-IQ3_S.gguf \
  --spec-type draft-mtp \
  --prompt "Write a short greeting." \
  -n 32 \
  -ngl 0

CPU smoke tests

• All eight tiers passed GGUF structural checks and text-generation smoke tests.
• The IQ3_S MTP draft-context test accepted 31 of 139 draft tokens (22.3%).
• The IQ3_S image test loaded the shared BF16 projector and read the headline “MEN WALK ON MOON” from the test image.

Provenance

sensenova/SenseNova-Loop-9B-Preview BF16 checkpoint
  revision 90dab7a3bf91922f5775be8deb6a290c128cfb90
  → full text+MTP BF16 GGUF, including block 56
  → CPU per-tensor quantization with an MTP-aware importance matrix
  → separate shared BF16 Qwen3VL mmproj

The converter is based on llama.cpp revision fad174ecd8bf22746a626c58e67e193f095aaa46 with local SenseNova-Loop support. Calibration uses the workspace's quantization_calibration.txt corpus; the importance matrix was generated with NextN/MTP collection enabled. “GSQ-RCO style” refers to this workspace's local per-tensor allocation workflow and is not a bit-exact reproduction of the upstream GSQ/RCO algorithms.

All conversion, quantization, and smoke testing are CPU-only. See the upstream model card for intended use and license details.