This standard quant uses the original SenseTime weights at fa381f7c466739475e172a7d3234e77351e4f613.
It retains a BF16 output head, the original native 524,288-token position encoding, and
144 FP32 recurrent and router tensors. The input embedding, vision linears and native draft
linears are 8-bit. Exact allocations and serialized storage are in PRECISION-ALLOCATION.csv,
TENSOR-INVENTORY.csv and quantization_config.json.
The release was checked with two independent GPU loads, text and vision fixtures, and four target-only streams using a shared 262,144-token Q8 cache. The sampled target-only peaks were 14.71 and 15.30 GiB in the two test shapes. The native speculative weights are included. Exact speculative parity and a physical 16 GiB deployment are separate, unfinished checks. Use target-only mode for the verified setting.
Load this directory with the EXL3 runtime in runtime/exllamav3-sensenova-source.tar.gz. It includes the SenseNova
Loop architecture port. Build the extension for your CUDA/PyTorch environment following
the included runtime README. The tested serving settings are in runtime/REFERENCE-PROFILE.json.
The max request length in that serving profile is bounded by its shared cache, while the
checkpoint's native position settings are unchanged.
model-00001-of-00002.safetensors 3.6 GBmodel-00002-of-00002.safetensors 2.7 GBsensenova/SenseNova-Loop-9B