OpenVDN/vdn-minimax-h3-edge

VDN-H3 Edge

Prepared FP8 weights and portable AdaLN constants for FreeVideo. Powered by MiniMax H3.

Two FP8 formats live in this repository. FreeVideo downloads only the matching format, about 22.9 GB / 21.4 GiB, plus the separately installed text encoder and decoders. It does not download both variants or the original transformer weights.

Formats

Directory	Quantization	FreeVideo automatic route
cache/	Per-tensor FP8	Blackwell, including RTX 50
rowwise/cache/	Per-channel FP8 weights	Ada / RTX 40 and Hopper / SM90; Ampere / RTX 30 uses this storage with BF16 compute

These routes preserve the engine's existing arithmetic. Ada also supports per-tensor FP8; keeping its current rowwise policy is a software/output compatibility choice. Windows rowwise uses FreeVideo's scalar FP8 GEMM plus fused row/column scaling, because the pinned PyTorch Windows build excludes its native rowwise implementation. Linux uses the native rowwise path. SM90's favorable rowwise dispatch does not establish equal performance on every architecture or shape.

The per-tensor model files are unchanged. Each format has its own immutable installer pin, so adding rowwise does not move or redownload an existing per-tensor installation. The repositories on Hugging Face and ModelScope contain the same payload. Setup selects a working source and resumes interrupted files.

Use

Install FreeVideo, open its launcher and start setup. The installer detects the GPU architecture and operating system, downloads the matching prepared files and checks the selected kernels on the machine. Users do not choose a quantization format or download both variants.

The model is public; no Hugging Face login is needed. Existing compatible local models are reused. The same automatic selection applies to FreeVideo's ComfyUI integration and command-line setup.

Portable AdaLN constants

Both formats contain the same two table sets: standard eight-step text-to-video and visual-keyframe schedules. I2VA, L2VA, FL2VA and visual-reference requests share the visual schedule where the engine's exact timestep contract matches. Prompt, seed, dimensions, duration, GPU model and Torch/CUDA versions do not require regenerating these stored constants.

Original AdaLN projections are omitted, saving approximately 26.02 GB per installation. Unsupported schedules or modulation-changing LoRAs can use the optional original projections from the pinned historical revision. Ordinary attention/FF LoRAs retain the tables.

Validation and provenance

The formats derive from the same 101 merged BF16 groups. Rowwise packaging copies all 363 previously prepared FP8 matrices without requantization; modification notices change headers only. Original projection identities and time-embedding tensors match the portable assets. All 50 text-only tables are byte-identical to the earlier rowwise generation cache.

The earlier rowwise native/Windows-compatibility comparison completed an eight-step 1344×768, 243-frame request on Linux SM120 with bit-identical video/audio latents, decoded RGB and audio. The existing per-tensor slim package was checked against the full prepared model on a complete 1344×768, 362-frame request, also bit-identically. This publication adds a byte-preserving rowwise package; it adds no new physical Hopper, Ada or Windows benchmark.

Individual cache format: freevideo-fp8-slim-v1; portable tables: freevideo-adaln-v2. These are FreeVideo engine packages. The source is OpenVDN/vdn-minimax-h3@751739ee5b9e3ac802dca5d5111075fdaeb47885; no additional training was performed. bundle.json, rowwise/cache/publication.json, SHA256SUMS and READY.json record source identities and publication verification. MODIFICATIONS.md describes the changes; NOTICE retains attribution.

License

VDN-H3 is a derivative of MiniMax-H3 and is distributed under the MiniMax H3 Community License Agreement, included here verbatim from the upstream repository in LICENSE.

The agreement grants rights only in its applicable territory, defined as worldwide excluding the European Union, the United Kingdom, the Republic of Korea, and the United States of America. It states that use outside the applicable territory is not authorized and invites people in an excluded territory to contact MiniMax about obtaining a license.

The agreement also contains redistribution requirements and an Acceptable Use Policy. Among other requirements, a distribution must include the agreement, modified files must carry notices of their modification, and distributions to third parties other than through hosted services must include the NOTICE file supplied with the code repository. Please read the agreement in full before using or distributing VDN-H3. This note is not a substitute for the license text or for legal advice.

OpenVDN's code license is provided separately in OPENVDN-CODE-LICENSE. It does not replace or alter the license governing the model weights and derived tables.