← sources

SenseNova-Loop-9B official GGUFs and a 4-bit AWQ are up (5.8 GB at Q4_K_M), plus their llama.cpp fork to run them

source: reddit (reddit:LocalLLaMA)published 2026-11-18 13:27ingested 2026-11-19 02:39

retained text

The official quants for SenseNova-Loop-9B (the 9B Looped Transformer from earlier this week) are here: https://huggingface.co/sensenova/SenseNova-Loop-9B-GGUF

The repo has Q4_K_M, Q5_K_M, Q8_0 and BF16 (sizes in the screenshot), plus a 0.9 GB mmproj for image input.

The sensenova_loop architecture isn't in mainline llama.cpp yet (issue #31571 is open), so for now these run on the fork they published, OpenSenseNova/llama.cpp. The GGUF card pins the commit, and the card's command once it's built is:

llama serve -hf sensenova/SenseNova-Loop-9B-GGUF:Q4_K_M -ngl 99 -c 4096 --port 8000

There's also a 4-bit AWQ, with a matching MTP draft model in the same repo: https://huggingface.co/sensenova/SenseNova-Loop-9B-AWQ

Their README at https://github.com/OpenSenseNova/SenseNova-Loop puts total GPU memory (4-bit weights plus an 8-bit KV cache) at about 10 GiB for 64K context and 15 GiB for 256K.

document fetched 2026-11-19 00:43 · canonical https://reddit.com/r/LocalLLaMA/comments/1xzf6im/sensenovaloop9b_official_ggufs_and_a_4bit_awq_are