Hugging Face's logo

igorls
/
Qwen3.8-Flash-Next-mixed-NInfer

Image-Text-to-Text
NInfer
qwen3.8
flash-next
nvfp4
fp8
int4
mixed-precision
blackwell
rtx-pro-6000
windows
single-gpu
multimodal
speculative-decoding
Model card Files
xet
 Community
Qwen3.8-Flash-Next-mixed-NInfer / NOTICE.md
igorls's picture
igorls
docs: publish Flash-Next NInfer preview metadata and source licenses
764886a verified
preview code
|
Raw
History Blame Contribute Delete
2.6 kB

Attribution and modifications

Qwen

Qwen3.8-Flash-Next is developed by Qwen. Copyright (c) 2026 Qwen. The original model and derived weights are subject to the Qwen Community License 1.0 included as LICENSE, copied without modification from:

https://huggingface.co/Qwen/Qwen3.8-Flash-Next/blob/de4b8e4d43b917e7706784d8bb445c9af86a3540/LICENSE

All six frontend resources embedded in this artifact are byte-identical to the official repository at that revision. This is a license/frontend reference; it does not identify the original BF16 weight revision used by the quantization publisher.

Primitive AI

The compute weights come from primitive-ai/Qwen3.8-Flash-Next-mixed-NVFP4-FP8 at a4e813ed3cfbbcc61e2929699eccb864a4dfa843. Primitive AI provides the main NVFP4 expert and FP8 projection quantizations and distributes the remaining BF16 components. Its model card specifies Qwen Community License 1.0.

The INT4 PLE tables come from primitive-ai/Qwen3.8-Flash-Next-PLE-quant at da8b39586016d8325ac619be28ad77d6296625ec, using only ples_int4. That repository's model card advertises Apache-2.0. It contains no separate LICENSE or NOTICE file at the pinned revision. LICENSE-APACHE-2.0.txt supplies the Apache license text alongside this attribution; those metadata do not relicense the underlying Qwen-derived weights or supersede Qwen's conditions.

NInfer conversion and runtime

NInfer upstream: https://github.com/Neroued/ninfer

NInfer workstation fork and this release: https://github.com/igorls/ninfer

Changes represented by this release are native .ninfer v2 packaging, projection concatenation/reordering, expert-bank aggregation and scale-layout conversion, substitution of Primitive's INT4 PLE table for the mixed source's BF16 PLE table, and embedding of the tokenizer/template/frontend resources. The release performs no fine-tuning or expert pruning. The 48 main MoE layers retain 512 experts each.

MTP expert banks remain BF16 in the distributed file. When MTP is enabled, the pinned NInfer loader converts these banks to its NVFP4 execution layout using a weights-only per-expert/per-group quantizer. This runtime conversion is NInfer's implementation, distinct from Primitive's main-expert source quantization. It does not change the published file. No optional pre-spliced NVFP4 MTP artifact is included in this release.

The historical converter commit was not recorded in the original conversion report. Exact source revisions, artifact bytes and SHA256, frontend hashes, and a compatible engine commit are recorded in artifact-manifest.json.