PixelML/Qwen3.8-Flash-Next-NVFP4-Dual-DGX-Spark

Qwen3.8-Flash-Next NVFP4 — Dual DGX Spark mirror

This is an exact, attributed mirror of RadixArk/Qwen3.8-Flash-Next-NVFP4 revision 7b719225242aacd3dbd3f9407468c2ee9a9d2594, published by PixelML so the checkpoint and its reproducible two-node NVIDIA DGX Spark deployment are easy to find together.

PixelML did not create or alter this quantization. RadixArk performed the quantization with NVIDIA Model Optimizer. PixelML independently validated the checkpoint, adapted SGLang for GB10/SM121, enabled NEXTN/MTP, measured the serving profile, and published the deployment recipe.

Reproducible deployment

Code, pinned container, SM121 compatibility patch, secure launcher, functional tests, and measured TPS:

PixelML/qwen3-8-flash-next-sglang-2x-dgx-spark

The validated configuration uses:

• Two NVIDIA DGX Spark systems, one GB10 GPU per node, connected with RoCE.
• SGLang tensor parallelism TP=2.
• NVFP4 routed experts through ModelOpt and FlashInfer CUTLASS GEMM/MoE.
• 262,144-token context and NEXTN/MTP speculative decoding.
• A pinned source revision and reproducible model verification.

PixelML measurements

Direct SGLang measurements with NEXTN/MTP and fixed 192-token coding requests:

Concurrent requests	Aggregate output tok/s
1	47.54
4	87.55
8	158.17
16	275.37

NEXTN/MTP improved repeatable single-stream throughput from 26.09 to 47.54 tok/s. These are deployment measurements, not model-quality benchmark scores. See the results/RESULTS-2026-08-26.md report for methodology and caveats.

Quantization provenance

Per the upstream RadixArk model card:

• Source model: Qwen/Qwen3.8-Flash-Next.
• NVIDIA Model Optimizer NVFP4 W4A4 quantization.
• Only routed experts in the 48 main-model MoE layers are NVFP4-quantized.
• Attention, QSA, GDN, mHC, shared experts, routers, embeddings, LM head, vision, and MTP tensors remain BF16; PLE n-gram tables use FP8 weights.
• Calibration used 128 cnn_dailymail articles with additional representativeness probing.

Refer to the RadixArk model card for its evaluation results, integrity evidence, and complete conversion notes.

Download

huggingface-cli download \
  PixelML/Qwen3.8-Flash-Next-NVFP4-Dual-DGX-Spark \
  --local-dir ./Qwen3.8-Flash-Next-NVFP4

For dual-Spark serving, use the linked GitHub recipe rather than an unpinned generic launch command.

License and attribution

The checkpoint is governed by the Qwen Community License 1.0, included in LICENSE. The original model is Copyright © 2026 Qwen. The NVFP4 conversion and its published validation evidence are attributed to RadixArk. PixelML's deployment code is separately available under the MIT License in the linked GitHub repository.

Users are responsible for reviewing the Qwen license, including its separate commercial-license conditions for Model-as-a-Service and AI Work Assistant businesses.