Swift-1.5-Qwen3.8-Flash-Next-NVFP4-CSF
Lossless format conversion of ukisai/Swift-1.5-Qwen3.8-Flash-Next-NVFP4 into the NVFP4-CSF checkpoint layout used by local-inference-lab Blackwell vLLM builds.
- Weights are unchanged. Only routed-expert block-scale tensors are re-encoded (lossless
compression:
weight_scale->*.nvfp4_csf_fixed+*.nvfp4_csf_exceptionsstreams). Every other tensor is byte-identical to the source checkpoint. - This layout lets a single 96 GB GPU serve the model: the CSF path keeps the large PLE n-gram table
host/disk-backed instead of GPU-resident, and enables the
nvfp4_csfkernels. - Serve with:
vllm serve <dir> --quantization nvfp4_csf --load-format nvfp4_csf(hardware support required, e.g. RTX PRO 6000 with local-inference-lab's vLLM). - License: see LICENSE (inherited from the source model).
- Downloads last month
- 38
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support
Model tree for deluxetiky/Swift-1.5-Qwen3.8-Flash-Next-NVFP4-CSF
Base model
Qwen/Qwen3.8-Flash-Next Finetuned
ukisai/Swift1.5-Qwen3.8-Flash-Next