edp1096/Huihui-RadixArk-Qwen3.8-Flash-Next-abliterated-NVFP4

Huihui-RadixArk-Qwen3.8-Flash-Next-abliterated-NVFP4

Community model combining RadixArk NVFP4 with changes extracted from Huihui’s abliterated GGUF. Distributed as safetensors, with RadixArk’s original MTP, vision weights and tokenizer retained.

Tested

• —TP1 / one DGX Spark: 64K context setting; basic text, code, tool-call and image checks passed.
• —TP2 / two DGX Sparks: 1M context setting with runtime YaRN ×4; basic checks and a 66K-token retrieval request passed. Full 1M input quality was not tested.

Changes were recovered from quantized GGUF weights, so this is not an exact Huihui BF16 reconstruction. Original activation scales were retained without recalibration. Comprehensive quality and refusal-removal benchmarks remain untested.

Source revisions and weight hashes

Sources and license

Qwen Community License 1.0. Credits: Qwen, RadixArk, Huihui, Unsloth, NVIDIA ModelOpt, and llama.cpp.