Will It Run AI
sand ai

MAGI-1

Frontier

by Sand AI

24B autoregressive diffusion model for streaming video generation. Produces high-quality cinematic video with strong temporal coherence. Requires 80GB+ VRAM for full inference. Apache 2.0 licensed.

  • 24B autoregressive diffusion — streaming video
  • Strong temporal coherence across long sequences
  • Apache 2.0 — fully open for commercial use
  • Requires 80GB+ VRAM (H100/A100 recommended)

Your hardware

Detecting...

Parameters24B
Max Resolution1280×720
Max Frames120
FPS24
Architecture3D-DIT
Licenseapache-2.0

Image Quality Benchmarks

Measured quality metrics for MAGI-1 outputs.

Human Preference Score88%

How often humans prefer this model's output (0-100%)

Aesthetic Score8.0

Visual quality and composition rating (5-9 scale)

This model requires 67+ GB VRAM for basic video generation. A GPU with 24GB+ VRAM is recommended.

VRAM by Scenario

VRAM estimates at FP16 and FP8 precision. FP8 uses ~40% less memory with minimal quality loss. Grade shows how well each GPU handles the generation workload.

FP16 (full precision)

ScenarioVRAMRTX 4090 24GBRTX 3060 12GBRTX 4060 8GBMacBook Pro M4 Pro 24GB
512×512 · 25 frames65.2 GBFFFF
768×512 · 25 frames67.3 GBFFFF
768×512 · 100 frames73.6 GBFFFF
1280×720 · 25 frames75.8 GBFFFF

FP8 (quantized — ~40% less VRAM)

ScenarioVRAMRTX 4090 24GBRTX 3060 12GBRTX 4060 8GBMacBook Pro M4 Pro 24GB
512×512 · 25 frames35.1 GBFFFF
768×512 · 25 frames37.2 GBFFFF
768×512 · 100 frames43.5 GBFFFF
1280×720 · 25 frames45.6 GBFFFF

Optimization Tips

Turbo / LCM distillation

Use distilled scheduler at 4-8 steps for faster iteration

Run with Python

Run with Python (diffusers)
from diffusers import DiffusionPipeline
import torch

pipe = DiffusionPipeline.from_pretrained(
    "sand-ai/MAGI-1",
    torch_dtype=torch.float16
)
pipe.to("cuda")

frames = pipe(
    prompt="your prompt here",
    num_inference_steps=50,
    guidance_scale=7.5,
    num_frames=120,
).frames[0]
# Save frames or export as video

Get started

Setup instructions for running MAGI-1 locally

1. Download the model

Get the checkpoint from HuggingFace

2. Place in:

ComfyUI/models/checkpoints/

3. Launch ComfyUI

python main.py
Note: Video generation requires video output nodes. Install ComfyUI-VideoHelperSuite from the ComfyUI Manager for SaveAnimatedWEBP or VHS_VideoCombine nodes.

Memory Breakdown

VRAM allocation for 25 frames at 768×512 on RTX 4090 24GB

Required: 67.3 GBAvailable: 24.0 GB
Weights48.0 GB
VAE0.2 GB
Text Encoder9.4 GB
Activations6.0 GB
Overhead0.5 GB

Estimated Generation Time

25 frames at 768×512, 30 steps, FP16.

RTX 4090 24GB~3m 55s
RTX 3060 12GB~14m 45s
RTX 4060 8GB~22m 15s
MacBook Pro M4 Pro 24GB~31m 38s

Sample Outputs

Available Formats & Downloads

Download MAGI-1 in different precisions. Lower precision = less VRAM but slight quality loss.

FormatoPrecisãoTamanhoProvedor
safetensorsRecomendadoBF1648.0 GBofficialBaixar

Related Workflows

You might also like

Frequently asked questions

FAQ — MAGI-1

How much VRAM does MAGI-1 need for video?

MAGI-1 (24B parameters) requires approximately 67.3 GB of VRAM at FP16 precision for generating 25 frames at 768×512. Video generation typically requires more VRAM than image generation due to temporal attention layers.

Can I run MAGI-1 on RTX 4090?

MAGI-1 exceeds the RTX 4090's 24 GB VRAM at FP16 for video generation. Consider reducing resolution, frame count, or using a GPU with more VRAM.

How long does it take to generate a video with MAGI-1?

On a reference GPU (RTX 4090 24GB), MAGI-1 generates a 25-frame video at 768×512 in approximately ~3m 55s at FP16 with 30 inference steps. Faster GPUs with higher memory bandwidth will reduce generation time.

What resolution and frame count does MAGI-1 support?

MAGI-1 supports up to 1280×720 resolution and 120 frames per generation at 24 FPS. Higher resolutions and frame counts require proportionally more VRAM.

About MAGI-1

Use cases
video-generationstreaming-videocinematic
Recommended runtimes
diffusers

See also