MiniMax-H3 Ref2VA (4-Bit INT4 Safetensors)

# 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 .safetensors (Single 9.80 GiB file)

- Quantization Scheme: Symmetric 4-Bit Linear FastInt4Linear) with FP16/BF16 Scale & Bias preservation

- AdaLN Conditioning: Pruned & Restored Continuous AdaLN Curve (1025 grid points)

- VRAM Footprint: 9.80 GiB (100% GPU Resident, Zero PCIe Swapping)

- Loading Time: ~0.03 seconds

## ⚡ Performance & Compatibility

- TeaCache Compatible: Full native support for Block-Level TeaCache (~50% compute bypass).

- LoRA / Adapter Support: Fully compatible with LightX2V 4-step Turbo LoRA minimax_h3_ref2v_turbo_4step_v0.1_bf16.safetensors).

- Attention Engines: Supports SageAttention 2.0 Patched, FlashAttention-2, and native PyTorch SDPA.

## 📜 License & Attribution

- Base model licensed under the MiniMax H3 Community License Agreement, Copyright © 2026 MiniMax.

- Powered by MiniMax H3.