MiniMax-H3 INT8 Release: Why Keep FC2 in BF16

marres · reddit · 2026-08-22

Released INT8 and INT8 ConvRot quantized versions of the MiniMax-H3 Pruned Ref-Delta Fused r1024 checkpoint for ComfyUI.

Quantization Strategy: Out of 50 transformer blocks, attn.qkv, attn.out, and mlp.fc1 (150 layers) are quantized to INT8, while mlp.fc2 is deliberately kept in BF16.

Technical Reason: A fully INT8 version failed on large sequences due to a specific issue in ComfyUI's fused linearinputact(..., "swiglu") path. This path dynamically quantizes the full activation matrix before the fc2 multiplication, attempting a 491 MiB contiguous INT8 scratch allocation that failed even with 47 GiB free VRAM. Keeping FC2 in BF16 avoids this path, allowing the workflow to complete.

Variants: The regular INT8 version uses native tensor-wise quantization. The INT8 ConvRot variant rotates weights before quantization (group size 256) to distribute outliers more evenly. Model size is approx 24.2 GB.

Original post →

More from Infra

Infra channel →