Release model card for FastH3 Trim

README.md

@@ -1,16 +1,23 @@
-# FastH3 Pruned 8-Step (full NVFP4, checkpoint 300)
+---
+license: other
+base_model: FastVideo/FastVideo-FastH3-8-Step-V2
+tags: [video-generation, text-to-video, audio, fastvideo, nvfp4, pruned]
+---
+# FastH3 Trim, 8-step, NVFP4, for Blackwell GPUs and DGX Spark
 
-**Full NVFP4 (W4A4), calibrated.** Every block's MLP (`ff.fc_in`, `ff.fc_out`), attention projections
-(`attn.to_q/k/v`, `attn.to_out`) and VSA compression gate (`attn.to_gate_compress`) are NVFP4, each with a **static
-activation scale max-calibrated on 1,000 prompts x all DMD steps**, stored in `transformer/nvfp4_weights.safetensors`
-as `_nvfp4_input_global_sf`. AdaLN, token refiner, embeddings and heads stay in their source precision.
-Requires an NVIDIA Blackwell GPU (sm100/sm120). This is the layout of the RTX 5090 speed configuration
-(the whole DiT stays resident on a 32 GB card).
+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.
 
-- 42-block pruned FastH3, rank-16 AdaLN, 8-step DMD (999..125), VSA 0.8, training checkpoint 300.
-- `text_encoder`: Qwen3-VL conditioner, NVFP4, 50 of 64 language layers (lossless: H3 reads hidden state 50).
-- `vae`: light VAE: 26-block LynnReal decoder + Kijai int8-convrot overlay.
-- `fastvideo_inference.json`: the sampling contract; FastVideo reads it automatically.
-- Load with `engine.quantization = {"transformer_quant": "NVFP4", "layer_profile": "h3_dit_vsa"}`.
+- **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:** 11.1 GiB. Attention, MLP and sparse-attention gate linears in NVFP4, each with a static activation scale calibrated on 1,000 prompts across all eight steps. For the RTX 5090, RTX PRO 6000 and DGX Spark.
+
+Full-quality counterpart in the same format: [FastVideo/FastVideo-FastH3-8-Step-V2-NVFP4-Consumer](https://huggingface.co/FastVideo/FastVideo-FastH3-8-Step-V2-NVFP4-Consumer).