Qwen3.8-27B-nvfp4-DFlash2-NInfer

Qwen3.8-27B DFlash 2 NVFP4 (NInfer)

DFlash-augmented NVFP4 image of Qwen3.8-27B for the NInfer engine: the registered qwen3.8-27b/nvfp4 fleet image (1124 objects) plus the 66-object dflash/ section (DFlash 2 drafter, 2,226,792,960 payload bytes). Published under the new nvfp4-dflash2 weights identity; the 27B target registry resolves it to the same nvfp4 weights profile, and the dflash group's residency is selected by the Engine startup features.

Identity

filename   = qwen3_8_27b_nvfp4.ninfer
model_id   = qwen3.8-27b
weights_id = nvfp4-dflash2
target_key = qwen3_8_27b
recipe_id  = qwen3_8_27b_nvfp4-v1
objects    = 1190 (1184 tensors + 6 resources)

Integrity

bytes   = 23719496192
sha256  = 6cc7560ae3427d8fa87b75c17e41328116b71b068c4c4dc06137fb73b656f64e

Sources

Note on the quantized source: the revision pinned by the artifact reference (60e813d4dbbdc5d64cf3f5a8caf2897bedf03679) no longer exists — the upstream repo was super-squashed on 2026-08-15 (16b6615af354), and the only commit since is a README-only update (7d6f8d4d). This build uses that main tree; the converter preflight validated the full 1118-tensor inventory, config facts, and dtypes against the source.

Build

Built by tools/convert/qwen3_8_27b/convert_nvfp4.py (--dflash-model), commit 87b4fe19 of taylor-shift/ninfer (branch dflash2-27b). The conversion report is stored alongside the artifact (qwen3_8_27b_nvfp4.ninfer.conversion.json).

Loading

Load with an NInfer 27B target build that accepts the nvfp4-dflash2 identity (DFlash 2 branch or later). The dflash startup feature selects the 66-object dflash residency; without it the image loads and runs as the plain nvfp4 model.