MiniMax-H3 / HyperFlow-H3-Flash-NOTICE

HyperFlow H3 Flash: low-rank down factors quantized to INT8 ConvRot (smaller files, faster steps); NOTICE updated

d1c34b0 verified 4 days ago

2.11 kB

MiniMax H3 is licensed under the MiniMax H3 Community License Agreement, Copyright © 2026 MiniMax. All Rights Reserved.

The HyperFlow H3 Flash weights in this repository are a Model Derivative of MiniMax-H3
(https://huggingface.co/MiniMaxAI/MiniMax-H3) as defined in Section I.11 of that Agreement. Video Rebirth created these
weights (Copyright 2026 Video Rebirth) and distributes them under the same MiniMax H3 Community License Agreement. A
copy of the Agreement is included in this repository as LICENSE.

Modifications from MiniMax H3 (Section III.2):
- fl2va_int8_convrot/ and ref2va_int8_convrot/: The transformer weights are pruned, low-rank factored, and quantized to
  rotated W8A8 INT8.
- fl2va_gguf/ and ref2va_gguf/: The compressed transformer weights are converted to the stable-diffusion.cpp tensor
  layout and quantized to GGUF.
- hyperflow_h3_flash_transformer.py: Modified from the diffusers MiniMax-H3 transformer (paired with
  hyperflow_h3_flash_int8.py for rotated INT8 linear layers).
- The upstream text encoder, VAEs, tokenizer, and processor are not redistributed here and remain the official
  MiniMax-H3 files.

HyperFlow H3 Flash is not an official MiniMax product and is not affiliated with or endorsed by MiniMax.
Powered by MiniMax H3.

---

Repacked for WanGP (2026): MiniMax-H3-FL2VA-HyperFlow-Flash-14B_int8_convrot.safetensors and
MiniMax-H3-Ref2VA-HyperFlow-Flash-14B_int8_convrot.safetensors are fl2va_int8_convrot/ and ref2va_int8_convrot/ of
https://huggingface.co/VideoRebirthAI/hyperflow-h3-flash (revision 9e9ed14b8be4b751d7517b8ca58dd199d509720c), each merged into
one file. Modifications: tensor keys renamed to WanGP's MiniMax H3 names; the two row halves of each feed-forward input
projection swapped (gate | value order); the INT8 weights and their scales stored as WanGP INT8 ConvRot tensors; the RoPE
frequency buffer added; the low-rank down factors, stored in BF16 upstream, quantized to INT8 ConvRot (per-row
scales, Hadamard groups of 256). Other tensor values are unchanged. The Agreement mentioned above is included here as HyperFlow-H3-Flash-LICENSE.

	MiniMax H3 is licensed under the MiniMax H3 Community License Agreement, Copyright © 2026 MiniMax. All Rights Reserved.

	The HyperFlow H3 Flash weights in this repository are a Model Derivative of MiniMax-H3
	(https://huggingface.co/MiniMaxAI/MiniMax-H3) as defined in Section I.11 of that Agreement. Video Rebirth created these
	weights (Copyright 2026 Video Rebirth) and distributes them under the same MiniMax H3 Community License Agreement. A
	copy of the Agreement is included in this repository as LICENSE.

	Modifications from MiniMax H3 (Section III.2):
	- fl2va_int8_convrot/ and ref2va_int8_convrot/: The transformer weights are pruned, low-rank factored, and quantized to
	rotated W8A8 INT8.
	- fl2va_gguf/ and ref2va_gguf/: The compressed transformer weights are converted to the stable-diffusion.cpp tensor
	layout and quantized to GGUF.
	- hyperflow_h3_flash_transformer.py: Modified from the diffusers MiniMax-H3 transformer (paired with
	hyperflow_h3_flash_int8.py for rotated INT8 linear layers).
	- The upstream text encoder, VAEs, tokenizer, and processor are not redistributed here and remain the official
	MiniMax-H3 files.

	HyperFlow H3 Flash is not an official MiniMax product and is not affiliated with or endorsed by MiniMax.
	Powered by MiniMax H3.

	---

	Repacked for WanGP (2026): MiniMax-H3-FL2VA-HyperFlow-Flash-14B_int8_convrot.safetensors and
	MiniMax-H3-Ref2VA-HyperFlow-Flash-14B_int8_convrot.safetensors are fl2va_int8_convrot/ and ref2va_int8_convrot/ of
	https://huggingface.co/VideoRebirthAI/hyperflow-h3-flash (revision 9e9ed14b8be4b751d7517b8ca58dd199d509720c), each merged into
	one file. Modifications: tensor keys renamed to WanGP's MiniMax H3 names; the two row halves of each feed-forward input
	projection swapped (gate \| value order); the INT8 weights and their scales stored as WanGP INT8 ConvRot tensors; the RoPE
	frequency buffer added; the low-rank down factors, stored in BF16 upstream, quantized to INT8 ConvRot (per-row
	scales, Hadamard groups of 256). Other tensor values are unchanged. The Agreement mentioned above is included here as HyperFlow-H3-Flash-LICENSE.