SuperQwen3.8-Flash-Next-abliterated-NVFP4-DGX-Spark

This is the measured NVFP4 DGX Spark deployment variant of Jiunsong/SuperQwen3.8-Flash-Next-abliterated@e7b3664cd9a5096a674db80f6ba3d379401cbc71. The derivative originates from Qwen/Qwen3.8-Flash-Next. It does not introduce a separate behavioral fine-tune; it converts the sealed SuperQwen payload using the immutable converter revision recorded below.

Variant

Format: NVFP4 DGX Spark
Quantization: {"algorithm": "NVFP4", "attention_and_nonexpert_modules": "BF16_or_inherited_FP8_as_sealed", "group_size": 16, "producer": "modelopt 0.46.0"}
Converter: NVIDIA/TensorRT-Model-Optimizer@87c9f8cf83021957d1a1a575c90c9a4eaaf7ef0c
Payload files: 229; payload bytes: 135260425234

Verified runtime

One real served inference call and the capability, tool-call, strict structured-output, reasoning, and formatting gates passed on 2x NVIDIA DGX Spark GB10, TP2 over direct RoCE v2 using sglang.

C1 decode: 18.098 tok/s
C1 prefill: 1015.614 tok/s
Directly tested context: 16431 tokens
Recipe: two-node SGLang with --tp-size 2 --nnodes 2; use the exact commands in release_evidence/runtime.json and replace only host-local paths

These numbers apply only to the sealed hardware, runtime, command, and artifact in release_evidence/runtime.json.

Integrity

The conversion and served runtime reports both bind to payload identity e83a9ac370e19c287eab43ebd786eea61027d2f8fc67693aac3104bf059001e0. SHA256SUMS.json seals every published file. Private-revision, anonymous public, and immutable-tag verification is recorded separately by the publisher.

Limitations

Measured only on 2x DGX Spark GB10 with SGLang TP2, EP1, and flashinfer_cutlass MoE.
The measured context gate is 16,384 tokens; no larger-context performance claim is made for this NVFP4 runtime.

This structural derivative changes refusal behavior relative to the upstream model. No compatibility, context, speed, quality, or universal behavior claim is made beyond the included measurements. Review the upstream Qwen license and terms before use.

Attribution

Upstream: Qwen/Qwen3.8-Flash-Next
Behavioral derivative: Jiunsong/SuperQwen3.8-Flash-Next-abliterated@e7b3664cd9a5096a674db80f6ba3d379401cbc71

SuperQwen3.8-Flash-Next-abliterated-NVFP4-DGX-Spark

Variant

Format: NVFP4 DGX Spark

Quantization:

{"algorithm": "NVFP4", "attention_and_nonexpert_modules": "BF16_or_inherited_FP8_as_sealed", "group_size": 16, "producer": "modelopt 0.46.0"}

Converter: NVIDIA/TensorRT-Model-Optimizer@87c9f8cf83021957d1a1a575c90c9a4eaaf7ef0c

Payload files: 229; payload bytes: 135260425234

Verified runtime

One real served inference call and the capability, tool-call, strict structured-output, reasoning, and formatting gates passed on 2x NVIDIA DGX Spark GB10, TP2 over direct RoCE v2 using sglang.

C1 decode: 18.098 tok/s

C1 prefill: 1015.614 tok/s

Directly tested context: 16431 tokens

Recipe: two-node SGLang with --tp-size 2 --nnodes 2; use the exact commands in release_evidence/runtime.json and replace only host-local paths

These numbers apply only to the sealed hardware, runtime, command, and artifact in release_evidence/runtime.json.

Limitations

Measured only on 2x DGX Spark GB10 with SGLang TP2, EP1, and flashinfer_cutlass MoE.

The measured context gate is 16,384 tokens; no larger-context performance claim is made for this NVFP4 runtime.