jmgao/Swift1.5-Flash-Next-mixed-NInfer

Swift 1.5 Qwen3.8-Flash-Next mixed quantization for NInfer

A quantization of ukisai/Swift1.5-Qwen3.8-Flash-Next using primitive-ai/Qwen3.8-Flash-Next-mixed-NVFP4-FP8's methodology to run on igorls's NInfer workstation fork on a single RTX Pro 6000.

The quantized PLE tables in Swift1.5 are identical to the base model, so we take ours from primitive-ai/Qwen3.8-Flash-Next-PLE-quant.

This slightly differs from igorls/Qwen3.8-Flash-Next-mixed-NInfer in that the MTP expert banks are stored in NVFP4 directly instead of BF16 and quantized at runtime.