license: other
license_name: qwen-community-1.0
license_link: https://huggingface.co/Qwen/Qwen3.8-Flash-Next/blob/de4b8e4d43b917e7706784d8bb445c9af86a3540/LICENSE
library_name: transformers
pipeline_tag: image-text-to-text
base_model:

• RadixArk/Qwen3.8-Flash-Next-NVFP4
• huihui-ai/Huihui-Qwen3.8-Flash-Next-abliterated-GGUF
  tags:
• qwen3.8
• nvfp4
• modelopt
• sglang
• abliterated
• gguf-delta-transfer

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

A community NVFP4 derivative by edp1096. This is not an official Huihui or RadixArk release. It transfers changes recovered from Huihui GGUF weights onto the RadixArk NVFP4 checkpoint while keeping the Hugging Face safetensors layout.

Construction

The complete Unsloth and Huihui UD-Q4_K_XL GGUF releases, including vision projectors, were compared tensor by tensor using SHA256. Of 1,558 tensors, 101 differ; all differing tensors use Q8_0:

• 48 shared-expert down projections;
• 36 Gated DeltaNet output projections;
• 12 full-attention output projections;
• routed-expert down projections in layers 2, 4, 30, 46, and 47 (512 experts each).

For the 96 dense/shared BF16 tensors, the transfer is:

BF16_original + dequant(Huihui_Q8) - dequant(Unsloth_Q8), rounded to BF16 with ties to even.

GDN output columns are restored from the GGUF tiled V-head layout to the HF grouped layout before applying the delta. The inverse follows llama.cpp commit 391fac16460f15233a7740550d858ac96df3419d.

For the five routed-expert layers, the same delta is applied to the official Qwen BF16 source weights, then quantized to NVFP4 using NVIDIA ModelOpt commit 87c9f8cf83021957d1a1a575c90c9a4eaaf7ef0c, block size 16, FP8 block scales and FP32 global scales. These tensors are not produced by dequantizing and requantizing the original NVFP4 weights.

Unchanged expert weights, PLE/ngram embeddings, vision weights, MTP weights and other non-target tensors retain RadixArk contents. The checkpoint contains 206 shards; only 24 shards differ from RadixArk. All 288,699 non-target safetensors entries were checked for equality. See PROVENANCE.json for pinned source revisions and shard hashes.

Limitations

• The delta contains Q8 quantization residuals. This is not an exact reconstruction of an unavailable Huihui BF16 checkpoint.
• Original RadixArk activation input scales are retained. No fresh activation calibration or general-purpose fine-tuning was performed.
• The name reflects the abliterated source; no comprehensive refusal-removal or safety benchmark was performed.
• Basic functional checks are not evidence of unchanged overall model quality.

Validation

On one DGX Spark, SGLang with NVFP4 routed experts and SSD-backed PLE/ngram embeddings successfully loaded the model with a 65,536-token context setting. Arithmetic, strict JSON, sorting, image color identification and a 6,037-token retrieval test passed. Generated Python code passed four execution cases; tool calls and Korean output were also checked.

For three 256-token generation trials with 41 input tokens, median client-stream throughput was 28.23 tok/s for RadixArk and 28.51 tok/s for this derivative. Median TTFT was 0.260 s and 0.258 s. The sample is small, generated text differs, and individual throughput varies; these values do not establish a speed advantage. No OOM or reboot occurred in the successful runs. These figures are for the isolated LLM runtime; co-hosting a diffusion model and other services requires a separate memory budget.

A two-DGX-Spark TP2 run also passed the basic checks with a 1,048,576-token server context setting, using a runtime YaRN factor-4 override relative to 262,144 tokens. The checkpoint configuration remains unchanged. A separate 66038-token retrieval request returned the correct key. This is not a full 1M-token input quality benchmark.

Use

Download with hf download edp1096/Huihui-RadixArk-Qwen3.8-Flash-Next-abliterated-NVFP4. Use a Qwen3.8 Flash Next compatible SGLang/vLLM implementation with ModelOpt NVFP4 support. DGX Spark deployments require an engine build with the corresponding SM121 kernels and PLE offload support; generic Transformers compatibility is not claimed by this card.

For non-thinking output, use chat_template_kwargs={"enable_thinking": false}. Keep the original tokenizer, processor and chat template distributed here.

Sources and license

The weights retain the Qwen Community License 1.0. Credit to Qwen, RadixArk, Huihui, Unsloth, NVIDIA ModelOpt, and llama.cpp.