MAGI-2 Preview: Sand.ai's Open-Source 114B Audio-Video Model
Sand.ai open-sources MAGI-2 Preview, a 114B-parameter MoE that turns text or images into 10-second videos with synchronized audio, activating only 6B parameters per token.
Sand.ai released MAGI-2 Preview on August 5, an open-source, 114B-parameter mixture-of-experts (MoE) model that generates 10-second videos with synchronized audio from a text prompt or a prompt plus a still image. The weights, inference code, and a technical report titled "Scaling Video Generation Models Efficiently" are all public.
Official MAGI-2 Preview visual from the Sand.ai announcement
MAGI-2 follows up on MAGI-1, Sand.ai's autoregressive video model from April 2025. MAGI-1 studied how video should be generated; MAGI-2 asks how a video generation model should scale, and the preview release validates that its architecture, training system, and data pipeline can scale together at the 100B level.
One Model for Video, Audio, and Text
MAGI-2 keeps a single-stream design: text, video, and audio tokens are concatenated into one unified sequence and processed by the same Transformer backbone. Instead of letting modalities interact only at narrow cross-attention interfaces, language, lip movement, body motion, sound, and camera rhythm exchange information throughout the model. The model generates both the visuals and the sound in one pass, then muxes the audio track into the output video file.
| Capability | Detail |
|---|---|
| Inputs | Text prompt (T2V), or text + still image (I2V) |
| Output | 10-second video with synchronized audio (the only supported duration) |
| Generation | Two-stage: preview stage at 512×896, refiner stage upscaling to 1088×1920 |
| Prompting | Long structured captions; the repo ships system prompts for LLM-based prompt enhancement |
Demo video from the official Sand.ai site
The official demo batch runs the same prompts as T2V and I2V entries, using stills like these as the image inputs:
 |  |
|---|---|
| Sample still from the official demo batch | Sample still from the official demo batch |
MagiMoE: 114B Parameters, ~6B Active per Token
The architecture combines a multi-head latent routing scheme with ultra-fine-grained experts to decouple total capacity from per-token compute. Each token's representation is split into 12 latent heads of 256 dimensions; every head has its own router and selects its top-6 experts from a pool of 256 narrow experts. Each sparse layer therefore contains 3,072 head-local expert units, of which only 72 are activated per token.
| Component | Configuration |
|---|---|
| Backbone | 40 Transformer layers (36 sparse, 4 dense boundary layers) |
| Model width | 3,072 |
| Routed representation | 12 heads × 256 dimensions |
| Expert pool | 256 experts per head, top-6 active |
| Expert FFN | 256 → 1,280 → 256, fused SwiGLU |
Architecture and systems overview from the MAGI-2 Preview technical report
Head Parallel keeps cross-device communication regular: activations are dispatched along the head dimension with statically known shapes, so buffers are preallocated and the dynamic routing is confined inside each head owner. The training system maps head-activation exchange onto InfiniBand across nodes and expert-state resharding onto NVLink within nodes, with the companion MagiCompiler and MagiAttention projects handling kernels and attention. On the data side, the report describes a shift from filtering-centric curation toward high-throughput data production with precise multimodal annotation.
Availability
Inference runs through the official Python pipeline (torchrun with a Docker image provided) and requires 8 NVIDIA Hopper GPUs — there are no ComfyUI custom nodes for MAGI-2 at this time. The complete checkpoint set is roughly 307 GB on Hugging Face: a 228 GB preview-stage transformer, a 14 GB refiner, the Qwen3.5-27B text encoder (56 GB), the Wan2.2 video VAE, the Stable Audio Open 1.0 audio VAE, and a distilled turbo VAE decoder used by default.
- Weights: sand-ai/MAGI-2-preview
- Code: SandAI-org/MAGI-2-preview
- Technical report: MAGI-2 Preview: Scaling Video Generation Models Efficiently
Comments
Sign in with GitHub to join the discussion.