Model card

README.md

@@ -9,8 +9,29 @@ MLX builds of [HyperFlow H3 Flash](https://huggingface.co/VideoRebirthAI/hyperfl
 
 | Precision | Files | Size |
 |---|---|---|
+| `4bit/` | 6 | 7.92 GB |
 | `6bit/` | 7 | 11.33 GB |
 | `8bit/` | 8 | 14.75 GB |
 | `bf16/` | 11 | 27.56 GB |
 
+
+## Builds
+
+The transformer only (`fl2va`: text and keyframe to video), converted from the publisher's `fl2va_int8_convrot`
+checkpoint: the int8 weights are dequantized and their Hadamard rotation undone, then stored as bf16 or MLX affine
+quantization (group size 64). No bf16 weights were ever released, so `bf16/` is the int8 model dequantized. Load with
+`mlx_model.py` from the Moonbox lab. The text encoder and VAEs are the base MiniMax-H3 ones.
+
+Checked block by block against the PyTorch model (each block fed the same input; pass = every block's cosine
+>= 0.995 and a step-0 video velocity within rel-L2 0.10 of the reference; the shipped W8A8 runtime is at 0.022):
+
+| Precision | Block rel-L2 median / max | Worst block cosine | Step-0 video rel-L2 | Status |
+|---|---|---|---|---|
+| `bf16/` | 0.0054 / 0.040 | 0.9992 | - | passes |
+| `8bit/` | 0.0081 / 0.043 | 0.9991 | 0.037 | passes |
+| `6bit/` | 0.0157 / 0.050 | 0.9987 | 0.052 | passes |
+| `4bit/` | 0.0584 / 0.160 | 0.9872 | 0.159 | **below threshold: experimental** |
+
+`4bit/` is published on request but fails the lab's threshold; expect visibly lower fidelity than `6bit/`.
+
 License: MiniMax H3 Community License (derivative; code Apache-2.0). These files carry the original model's license; read it before using them.