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.
More from Infra
- Data center water use: 10GW hyperscaler consumes 120B gallons annually — SumitGup · 2026-08-22
- OpenBot: Open-source AI Coworkers with Isolated Containers and Policy Gateways — aigclink · 2026-08-22
- Alibaba Cloud MaaS integrates GLM-5.3 and DeepSeek-V4-Pro — 创业邦 · 2026-08-22
- Nuclear 'hot rock' generates immense energy vs weak passive solar needing maintenance — tawnniee · 2026-08-22
- 2-4K GPUs can serve 100T tokens daily, sparking efficiency debate — teortaxesTex · 2026-08-22
- Prediction market opens on whether Apple will announce 1TB+ unified memory chip — benfielding · 2026-08-22