# MiniMax-H3 Ref2VA (4-Bit INT4 Safetensors) - Experimental Model quant, may have zombie sway and artifacts in 10 second runs. Pure int4 version, meant for lowest possible vram usage for experimentation. This repository provides an optimized 4-bit INT4 Safetensors quantization of MiniMax-H3-Ref2VA, engineered for consumer GPUs (16 GB VRAM, RTX 4080 / RTX 4090 / Laptop GPUs).## 🚀 Model Details- Base Architecture: MiniMax-H3 Ref2VA Diffusion Transformer (DiT)- Format: Zero-Copy <code>.safetensors</code> (Single 9.80 GiB file)- Quantization Scheme: Symmetric 4-Bit Linear <code>FastInt4Linear</code>) with FP16/BF16 Scale &amp; Bias preservation- AdaLN Conditioning: Pruned &amp; Restored Continuous AdaLN Curve (1025 grid points)- VRAM Footprint: 9.80 GiB (100% GPU Resident, Zero PCIe Swapping)- Loading Time: ~0.03 seconds## ⚡ Performance &amp; Compatibility- TeaCache Compatible: Full native support for Block-Level TeaCache (~50% compute bypass).- LoRA / Adapter Support: Fully compatible with LightX2V 4-step Turbo LoRA <code>minimax_h3_ref2v_turbo_4step_v0.1_bf16.safetensors</code>).- Attention Engines: Supports <code>SageAttention 2.0 Patched</code>, <code>FlashAttention-2</code>, and native <code>PyTorch SDPA</code>.## 📜 License &amp; Attribution- Base model licensed under the MiniMax H3 Community License Agreement, Copyright © 2026 MiniMax.- Powered by MiniMax H3.

Files with SHA256: f302a1612f5a76282dbd38f05da25db3e218c5ed8e742c66bf6aa78ef805fb95