Release model card for FastH3 Trim

README.md

@@ -1,12 +1,23 @@
-# FastH3 Pruned 8-Step (bf16, checkpoint 300)
+---
+license: other
+base_model: FastVideo/FastVideo-FastH3-8-Step-V2
+tags: [video-generation, text-to-video, audio, fastvideo, bf16, pruned]
+---
+# FastH3 Trim, 8-step, BF16 source weights
 
-Reference bf16 transformer; source for the FP8/NVFP4 exports.
+FastH3 Trim is an experimental, smaller version of [FastH3 V2](https://huggingface.co/FastVideo/FastVideo-FastH3-8-Step-V2):
+eight-step text-to-video with synchronized audio from 42 of the 50 MiniMax H3 transformer blocks. We removed the eight
+blocks whose removal changed video and audio predictions the least, replaced each block's AdaLN timestep projection with
+a shared rank-16 basis, and trained the result with eight-step DMD2 (checkpoint 300). Removing blocks makes the model
+faster and smaller but costs some quality; use FastH3 V2 when quality matters most.
 
-## Model
-- MiniMax-H3 student pruned from 50 to **42 transformer blocks** (removed blocks 6, 7, 9, 13, 15, 16, 22, 23), rank-16 AdaLN,
-  trained with 8-step DMD (timesteps 999, 874, 749, 624, 500, 375, 250, 125) and VSA sparse attention (sparsity 0.8, 64-token tiles).
-- Training checkpoint 300 of run `s42-r16-dmd8-vsa80` (selected over 800 for prompt adherence on a multi-seed evaluation).
-- Sampling contract: video/audio scheduler shift 10/3, guidance 1.0, VSA_sparsity 0.8, VSA_tile_size 64.
-- `text_encoder` is not included; it is identical to the one in `FastVideo/FastVideo-FastH3-8-Step-V2`.
+- **Sampling:** 8 DMD steps (999, 874, 749, 624, 500, 375, 250, 125), sparse attention keeping 20% of tiles.
+  `fastvideo_inference.json` holds the schedule, which FastVideo reads automatically.
+- **Text encoder:** NVFP4 Qwen3-VL trimmed to the 50 layers H3 reads.
+- **VAE:** LynnReal lightweight video VAE with Kijai's INT8 weights; H3 audio VAE.
 
-**Status: internal, private evaluation build.** License: MiniMax H3 Community License (inherited).
+Blog post: [FastH3 on Consumer Hardware](https://haoailab.com/blogs/fasth3-rtx/) · Code: [FastVideo](https://github.com/hao-ai-lab/FastVideo)
+
+- **Transformer:** 34.9 GiB in BF16 (FP16 AdaLN factors). Source for the quantized releases and for local MLX conversion on Apple Silicon.
+
+Full-quality counterpart in the same format: [FastVideo/FastVideo-FastH3-8-Step-V2](https://huggingface.co/FastVideo/FastVideo-FastH3-8-Step-V2).