MAGI-2¶

Development line: project:magi-2 · thread magi-2-preview
Last event: 2026-08-07 · 1 dated since 2026-08-07 · Researched: 2026-09-05 · confidence: high

What it is¶

MAGI-2 Preview is a 114B-parameter unified audio-video MoE model for teams generating 10-second clips with synchronized audio locally from text or images.

• T2V creates video from text prompts with joint audio.
• I2V animates a starting frame into video with joint audio.
• Two-stage pipeline generates a preview and refines output up to 1080p.

The model uses 6B active parameters per token, roughly 307 GB of weights, and eight Hopper GPUs. This is an infrastructure-heavy open preview release, not a lightweight API tool.

Development line¶

• 2026-08-07 — Sand AI published MAGI-2 preview materials. The model uses 114B parameters with about 6B active per token on a unified MagiMoE audio-video architecture. Repository documentation limits the release to 10-second clips, a preview stage with a refiner up to 1080p, and T2V and I2V with synchronized sound.

What changed¶

2026-08-07 — MAGI-2 Preview became available with weights and inference code. Primary material from Sand AI on 2026-08-05 specifies the scale: 114B parameters, about 6B active per token, and a unified MagiMoE audio-video architecture. Repository documentation limits the release to 10-second clips, a preview stage and refiner up to 1080p, and T2V and I2V with sound.

How to use this¶

As of 2026-08-07, practitioners should treat MAGI-2 as a publicly previewed Sand AI project and consult its linked repository, model page, and API v2 documentation before evaluating or integrating it; specific capabilities and access conditions still require direct verification.

1. Prepare a Linux environment with Python 3.12, current CUDA, ffmpeg, and eight NVIDIA Hopper GPUs as required by the inference code. — https://github.com/SandAI-org/MAGI-2-preview
2. Pull the Docker image or build from source. Specify the image tag matching the commit rather than moving latest for reproducibility. — https://github.com/SandAI-org/MAGI-2-preview
3. Download the full checkpoint set from sand-ai/MAGI-2-preview into the ckpt folder without renaming directories. — https://huggingface.co/sand-ai/MAGI-2-preview
4. Run T2V with a text prompt or I2V with a prompt and first frame. The output is a 10-second clip, and 1080p passes through preview and refiner. — https://github.com/SandAI-org/MAGI-2-preview

Best practices¶

• Pin the Docker tag to a specific commit and log it with results, because latest changes. — https://github.com/SandAI-org/MAGI-2-preview
• Keep the default staged offload for preview and refiner at 1080p, because both components and activations do not fit together even on an 80 GB GPU. — https://github.com/SandAI-org/MAGI-2-preview
• Use structured prompt expansion for the full 10-second scene when a manual prompt does not set action, framing, and audio. — https://github.com/SandAI-org/MAGI-2-preview

Superseded by this¶

• 2026-08-05 — MAGI-2 Preview is not replaced by a separate publicly dated subsequent release in verified primary sources; documentation marks the distilled variant as coming soon.
• 2026-08-07 — The assumption that MAGI-2 is a ready public API product is obsolete: the verified implementation provides open weights and inference code with cluster requirements.

Still unknown¶

• URL https://platform.sand.ai/docs/api-v2 returned no readable documentation during verification, so API features, pricing, limits, and current availability remain unconfirmed.
• Verified primary sources provide no separately dated event after 2026-08-07; changes in the current README cannot be reliably dated as a new release.

Sources¶

Agent brief¶

• Subject: project:magi-2, thread magi-2-preview, 1 dated events 2026-08-07 → 2026-08-07.
• Practical note: As of 2026-08-07, practitioners should treat MAGI-2 as a publicly previewed Sand AI project and consult its linked repository, model page, and API v2 documentation before evaluating or integrating it; specific capabilities and access conditions still require direct verification.
• Confidence: high. Dated supersedes above are the authority for what is obsolete.