base_model:
- nvidia/Qwen3.8-Flash-Next-FP4-Local
- huihui-ai/Huihui-Qwen3.8-Flash-Next-abliterated-v2-GGUF
pipeline_tag: image-text-to-text
library_name: vllm
license: other
license_name: nvidia-open-model-agreement
license_link: https://www.nvidia.com/en-us/agreements/enterprise-software/nvidia-open-model-agreement/
tags: [nvfp4, mxfp8, modelopt, abliterated, qwen3.8]
Huihui-Qwen3.8-Flash-Next-abliterated-FP4-Local
NVIDIA FP4-Local with the audited Huihui v2 changes transferred from matched Unsloth/Huihui Q8_0 weights. Changed projections are reconstructed, updated, rounded to BF16 and encoded as MXFP8. Routed NVFP4 experts and their CSF scales, PLE, MTP, vision and tokenizer retain the NVIDIA checkpoint bytes.
See transfer-manifest.json for pinned sources, changed tensors and quantization errors, and source-audit.json for the complete source comparison. Runtime results are recorded separately in runtime-qualification.json when available.
Sources: NVIDIA FP4-Local, Huihui v2, Unsloth Q8. The NVIDIA Open Model Agreement and the base model's Qwen Community License apply.
Local GPU qualification
The checkpoint was tested on DGX Spark with the pinned B12X runtime and factor-4 YaRN overlay recorded in runtime-qualification.json. Both the NVIDIA baseline and this checkpoint completed the actual 1,048,056-token two-field retrieval test and the following short request. Code execution, structured JSON, tool calls and the image smoke test passed.
One candidate check failed: the reasoning answer was correctly 445, but the final response included equations despite an integer-only instruction. Accordingly, the manifest keeps runtime_validated: false and records runtime_tested: true; the result is completed with findings, not an all-checks pass. Both variants also show prose-quality limitations in the limited samples and use additional swap during startup.
The full comparison and generated responses are in runtime-qualification.json and qualification/. The reproducible runtime patches are in qualification/runtime/. These measurements do not establish broad benchmark parity or donor refusal-rate equivalence. This is quantized delta transfer, with no new QAD training or activation calibration.