Great!
Isn't the quantum too small?
Artifact size note: This build is not directly size-comparable to the full upstream Qwen3.8 NInfer/NVFP4 artifacts. It was converted with --components text and currently does not bundle Vision, MTP or DFlash2 companion weights. D48 also retains NInfer's compact groupwise allocation for most tensors and selectively replaces only 48 MLP Down projections with NVFP4. As a result, the optimized artifact is smaller than its own 17.11 GB text-only NInfer baseline.
The smaller size does not imply that this build uses a more aggressive full-model NVFP4 profile. It is a selective mixed-precision configuration optimized around the measured ThinkingCap sensitivity map.
Isn't the quantum too small?
Since all of this works together, the draft modules don't come until the end
Quick update: my baseline is built with NInfer’s standard qwen3_8_27b recipe using --components text, so it is already a compact mixed Q4/Q5/Q6/Q8 build at 17.11 GB.
D48 keeps that baseline and only moves 48 selected MLP Down projections to NVFP4 based on a full 64-layer sensitivity sweep.
Baseline: 17,106,573,056 B, PPL 4.351060
D48: 16,705,529,348 B, PPL 4.348019
So the smaller size comes mainly from the text-only scope + NInfer’s compact groupwise recipe, not from a custom converter.
Next step is a ThinkingCap-specific IMatrix, then Gate/Up optimization.