Sol-H3-VAE Fast Video Decoder Model: 8GB+ VRAM, File List

At a Glance
| Item | Value |
|---|---|
| Repository | Efficient-Large-Model/Sol-H3-VAE |
| Publisher guide | Efficient Large Model (NVIDIA and MIT): models and licenses |
| Published | 2026-10-17 |
| Formats | safetensors |
| Source type | Primary source (the publisher itself) |
Values determined by this site’s code when the information was collected. Dates are JST.
Overview
Efficient-Large-Model has released “Sol-H3-VAE", a fast video decoder for MiniMax-H3. This model replaces the 36-layer H3 VAE decoder employed in MiniMax-H3 with a 24-layer noncausal decoder built through distillation.
This model takes normalized latents as input and outputs the exact same frames as the original H3 VAE decoder. Note that no changes have been made to the audio decoder.
Specifications
The main specifications and operating environment details described in the documentation are as follows:
- Architecture: 24-layer noncausal decoder (distilled from the H3 VAE decoder)
- Parameters: 1.62B
- Decoding method: Processes entire frames using windowed attention instead of spatial tiles, decoding each native 5-latent block bidirectionally
- Input specifications: Normalized latents from MiniMax-H3
- Output specifications: Identical frames to the H3 VAE decoder
- Required libraries and environment: PyTorch (CUDA-compatible), Triton, Diffusers (requires MiniMax-H3 support), FlashAttention-4 (
flash_attn.cute), QuACK - Test environment example: RTX 5090, PyTorch 2.12 (CUDA 13), Triton 3.7, Diffusers 0.40, FlashAttention-4 4.0.0b19, QuACK 0.6.3
Note that on the initial call, kernel compilation and CUDA graph capture are performed according to the shape of each latent (taking approximately 25 seconds in an execution example on an RTX 5090).
Performance and Quality
Based on the benchmark results provided in the model card, we compare it with the existing H3 VAE decoder.
| One 5-second clip | H3 VAE | Sol-H3-VAE |
|---|---|---|
| 768p, RTX 5090 | 12.5 s | 5.0 s |
| 720p, GB200 | 6.16 s | 0.34 s |
| 2K, GB200 | 14.5 s | 0.89 s |
From the comparison table above, it can be seen that Sol-H3-VAE significantly outperforms H3 VAE in decoding speed. For 720p decoding in a GB200 environment, Sol-H3-VAE is extremely fast at 0.34 seconds compared to 6.16 seconds for H3 VAE. Even at 2K resolution, a dramatic reduction from 14.5 seconds to 0.89 seconds is reported. Testing at 768p using an RTX 5090 also shows speedups from 12.5 seconds to 5.0 seconds.
As an indicator of image quality, PSNR (Peak Signal-to-Noise Ratio) is reported. Values of 45.4 dB are shown for held-out latents (720p to 2K) and 38.2 dB for generated 768p clips.
However, because this model is distilled, its quality is very close to but not entirely identical to the H3 VAE decoder. Furthermore, due to its noncausal design, it cannot decode latents one by one while streaming; the entire clip must be decoded at once. As a result, memory usage increases as the length of the clip increases.
Strengths and Use Cases
This model is intended for high-efficiency video decoding for MiniMax-H3. It is designed to be integrated as part of the Sol-H3 in ComfyUI pipeline, adopting an approach that uses windowed attention to process entire frames bidirectionally all at once. Being a distilled model, it is suited for cases where fast processing is desired while maintaining the quality of the original H3 VAE decoder.
Hardware Requirements
Estimated requirements (calculated by Local Model Watch)
| Your VRAM | Quantization | File size | Est. memory needed |
|---|---|---|---|
| 8GB (RTX 4060 / 3060 Ti, etc.) | Original precision | 6.0GB | 7.2GB |
Memory estimates add a 20% runtime overhead (KV cache, etc.) to the actual size of the distributed files. Actual usage varies with context length, batch size and inference engine. These figures are computed by this site from file sizes, not published by the model’s authors. Compare with other models in our VRAM quick reference. What the quantization names mean: glossary.
Our Own Measurements
Measurements We Did Not Take
This model does not state a license, so we cannot confirm permission to use it. Because this site carries advertising, we did not run the model (no CPU run, answers, quantization comparison, conversion or generation).
Distributed Files
Weight files published in Efficient-Large-Model/Sol-H3-VAE, listed by this site from the Hugging Face API. Sizes are the actual file sizes.
| File | Size |
|---|---|
fast_h3_vae_noncausal_24l.safetensors |
6.47GB |
How to Get It
It can be obtained from Hugging Face’s Efficient-Large-Model/Sol-H3-VAE via snapshot download or similar methods. Code repositories and specific implementation examples can be found on related pages for Sol-H3 in ComfyUI.
import sys
from huggingface_hub import hf_hub_download, snapshot_download
repo = snapshot_download("Efficient-Large-Model/Sol-H3-VAE")
sys.path.insert(0, f"{repo}/code")
from spark_decoder import SparkFastH3VAEDecoder
vae_config = hf_hub_download("MiniMaxAI/MiniMax-H3", "vae/config.json")
decoder = SparkFastH3VAEDecoder(vae_config, f"{repo}/fast_h3_vae_noncausal_24l.safetensors")
Using the model requires the H3 VAE config.json, but the H3 VAE weights themselves are not required.
Related Articles
- StreamRender-H3 Real-Time Video Generation Model: 8GB+ VRAM, File List
- H3-to-LTX-Latent-Adapter: 4GB+ VRAM, File List
- LongLive2.0-Robot-L Robot Video Generation Model: File List
- WORL-Helios Video Generation Model: File List
What to Read Next
- Find models by VRAM (This model runs from the 8GB tier) → Other models that run on a 8GB GPU
- Formats this model is available in → Safetensors format guide and models
- Learn about the publisher → Efficient Large Model (NVIDIA and MIT): models, licenses and articles

