MAGI-2 Preview is Sand.ai’s bet on a different scaling path for video: a 114-billion-parameter mixture-of-experts model that activates just 6 billion parameters per token. Released August 5, 2026 under Apache 2.0, it generates 10-second video clips with synchronized audio in a single pass.
This is a research preview, not a production product. Sand.ai is explicit about that — they call it “an intermediate validation” of their scaling approach, with controlled ablations and capability analyses still coming. But the fact that a 114B video MoE with full weights, code, and a technical report is freely available under Apache 2.0 makes it significant regardless.
Architecture
| Spec | Value |
|---|---|
| Total parameters | 114B |
| Active per token | ~6B (~5.3% of network) |
| Architecture | MagiMoE — ultra-fine-grained MoE with multi-head latent routing |
| Routing | 12 latent heads × 256 dims, top-6 selection |
| Text encoder | Qwen3.5-27B (56 GB) |
| Video VAE | Wan2.2 |
| Audio VAE | Stable Audio Open 1.0 |
Output specs
| Value | |
|---|---|
| Duration | 10 seconds (fixed) |
| Generation | Two-stage: preview at 512×896, refiner to 1088×1920 (1080p) |
| Audio | Synchronized, generated alongside video |
| Modes | Text-to-video (T2V), Image+text-to-video (I2V) |
Hardware requirements
| Component | Size |
|---|---|
| Preview transformer | 228 GB |
| Refiner | 14 GB |
| Text encoder (Qwen3.5-27B) | 56 GB |
| VAEs + turbo decoder | ~10 GB |
| Total checkpoint | ~307 GB |
Inference requires 8 NVIDIA Hopper GPUs via torchrun. No ComfyUI nodes or consumer-grade options exist at launch.
How it compares
- vs MiniMax H3: Both generate video with native audio. H3 is 33B dense, runs on an RTX 3060, and supports 15 seconds. MAGI-2 is 114B MoE, needs 8 H100s, and generates 10 seconds. H3 is more practical; MAGI-2 is more research-forward.
- vs FLUX 3 Video: FLUX 3 generates up to 20 seconds at Full HD with native audio, ranks higher on quality (AA #1 on their own evaluation). FLUX 3 is a commercial product; MAGI-2 is an open research release.
- vs Wan 2.7: Wan 2.7 is the established open-weight video model from Alibaba. MAGI-2 explores MoE scaling at a larger parameter count but with shorter output and higher hardware requirements.
Why it matters
MAGI-2 is the first 100B-scale video generation model released as a complete open-source package under Apache 2.0. The research contribution — proving that multi-head latent routing can decouple total capacity from per-token compute at this scale — is the point. A distilled version with fewer denoising steps is listed as coming soon, which could make it more practical.
Limitations
- Fixed 10-second duration. No variable length generation.
- 8 Hopper GPUs minimum. Not accessible for individuals or small teams.
- Research preview quality. AA Video Arena #6 — behind commercial products on quality benchmarks.
- No hosted API. Self-hosting only, with significant infrastructure requirements.
For the full video model landscape, see the video models directory.
Best for
- Research into efficient video generation scaling via MoE architectures
- Self-hosted video generation for teams with Hopper-class GPU clusters
- Fine-tuning and derivative model development under Apache 2.0
- Studying video MoE routing behavior at the 100B scale
Pros & cons
Strengths
- Apache 2.0 — fully permissive commercial license for a video generation model
- 114B total / 6B active — efficient MoE architecture scales capacity without proportional compute cost
- Native synchronized audio — video and sound generated in one pass, no separate audio pipeline
- Complete research release — weights, inference code, and technical report all public
Limitations
- Requires 8 NVIDIA Hopper GPUs — not accessible for most individual developers
- Fixed 10-second duration — no variable clip length
- Research preview — Sand.ai explicitly frames this as "an intermediate validation," not a finished product
- AA Video Arena
Frequently asked questions
When was MAGI-2 Preview released?
August 5, 2026, with weights on Hugging Face and inference code on GitHub.
What hardware does MAGI-2 require?
Inference runs through torchrun and requires 8 NVIDIA Hopper GPUs (H100 or equivalent). The complete checkpoint is roughly 307 GB across the preview transformer (228 GB), refiner (14 GB), Qwen3.5-27B text encoder (56 GB), Wan2.2 video VAE, and Stable Audio Open 1.0 audio VAE.
How does MAGI-2 generate audio?
Audio and video are generated in a single stream using a unified Transformer architecture. The audio latents are decoded by the Stable Audio Open 1.0 VAE and muxed into the output file. No separate TTS or audio model needed.
What is MagiMoE?
MAGI-2's custom MoE architecture. It uses multi-head latent routing with ultra-fine-grained experts to decouple total model capacity from per-token compute. Each token representation is split into 12 latent heads of 256 dimensions, with top-6 routing — activating ~6B of the 114B total parameters.
Is a distilled/faster version coming?
Sand.ai lists a 'MAGI-2 Preview, distilled' variant as 'coming soon' on the GitHub repo — promising 'far fewer' denoising steps.