SenseNova-Loop-9B-Preview-GGUF

GGUF quantizations of sensenova/SenseNova-Loop-9B-Preview — a looped multimodal model (looped decoder + manifold-constrained hyper-connections) from SenseNova. By AtomicChat.

[!IMPORTANT] New architecture — needs a patched llama.cpp. SenseNova-Loop is a sensenova_loop model: a Qwen3.5-class text backbone whose layers 4-27 are executed twice (56 layer computations from 32 stored layers) with two-stream manifold-constrained hyper-connections (mHC, Sinkhorn routing). Upstream llama.cpp, LM Studio and Ollama do not support this architecture yet, so these files will not load on a stock build. They 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). These are text-only GGUFs for now; the vision tower (mmproj) is a follow-up.

Files

Baseline preset quants (no importance matrix, no layout tuning) for a fast first release; improved quants to follow. The mHC routers are kept in F32 (routing is precision-sensitive).

File	bits/weight	Size on disk	Role
`SenseNova-Loop-9B-Preview-BF16.gguf`	16.0	16.7 GB	reference / highest fidelity
`SenseNova-Loop-9B-Preview-Q8_0.gguf`	~8.5	8.9 GB	near-lossless, fits comfortably in 16 GB
`SenseNova-Loop-9B-Preview-Q6_K.gguf`	~6.6	6.9 GB	high quality
`SenseNova-Loop-9B-Preview-Q5_K_M.gguf`	~5.7	6.0 GB	balanced
`SenseNova-Loop-9B-Preview-Q4_K_M.gguf`	~4.9	5.2 GB	smallest, runs on ~8 GB-class GPUs

Architecture notes

Text backbone is Qwen3.5-class: hybrid attention (three Gated-DeltaNet linear-attention layers per one full-attention layer), QK-norm, attention output gate, interleaved M-RoPE with partial_rotary_factor 0.25, dense SwiGLU FFN, vocab 248320.
Looped schedule: loop_start=4, loop_end=28, executed twice; each visit keeps its own attention KV and GDN recurrent state.
mHC: two parallel residual streams over the looped region, with a dense router (sigmoid pre/post gates + a Sinkhorn-normalized doubly-stochastic mix) replacing the plain residual add. Router weights are F32.
MTP draft block and the vision tower are not included in these text GGUFs.

Running

Build a llama.cpp with SenseNova-Loop support, then:

llama-cli -m SenseNova-Loop-9B-Preview-Q6_K.gguf -ngl 99 -c 4096 -st \
  -p "What is the capital of France? Answer in one short sentence."

The chat template (with thinking) is embedded in the GGUF. Verified greedy output on BF16: coherent reasoning, code and multilingual generations (e.g. the train-speed word problem resolves to 80 km/h; "capital of France" -> Paris).

Atomic Chat

Atomic Chat support: in progress (the engine needs this architecture; port underway).

Credits & license