Sand.ai

Sand.ai

3.0

MAGI-2: Unified Audio-Video Generation Model — One model jointly generates visuals and sound, with lip sync and cinematic camera movement, activating just 6B parameters per token.

Last Updated Nov 10, 2026
Sand.ai screenshot

About

Sand.ai is a foundation model company focused on audio and video generation. Its flagship model, MAGI-2 Preview, is a unified 114B-parameter audio-video generation model. Built on the company's proprietary MagiMoE architecture, it activates only 6B parameters per token and generates visuals and sound jointly from a single model, exploring an efficient path to scaling video generation. Its predecessor, MAGI-1, was released as an open-source autoregressive video generation model, and Sand.ai has continued that open-source tradition alongside its newer releases.

The core idea behind MAGI-2 is that a single model should handle both picture and sound at the same time, rather than generating video first and attaching audio afterward. This joint generation approach addresses a long-standing alignment problem in generated content, where lip movements, sound effects, and ambient noise can drift out of sync with the visuals. Alongside the model itself, Sand.ai operates a developer platform at platform.sand.ai, where API documentation is available for teams that want to integrate video generation into their own products and workflows.

Key Features

Unified audio-video generation: One model produces visuals and sound together, so dialogue matches lip movements, action matches sound effects, and environmental changes are matched by corresponding audio feedback.
Character performance: Generates dialogue, singing, and emotional shifts, with lip shape, facial expression, gaze, and body language coordinated naturally.
Cinematic camera movement: The model understands framing, composition, perspective, and camera intent, producing push-ins, tracking shots, and orbiting movements with pacing that resembles live filming.
MagiMoE sparse architecture: A mixture-of-experts design activates only 6B of the model's 114B total parameters per token, keeping inference costs close to those of a much smaller model.
Autoregressive generation: The model's autoregressive design naturally supports video continuation and interactive extension, letting users extend or build on generated clips.
Open-source lineage: Building on MAGI-1, an open-source autoregressive video generation model, Sand.ai continues to release work in the open-source ecosystem.
Developer API: The platform at platform.sand.ai provides documentation for integrating video generation capabilities via API.

Pricing & Fees

  • Pricing: Not publicly listed on the pages we reviewed; check the official site for current plans.
  • Single-model alignment: Joint generation of audio and video from one architecture removes the mismatch that comes from generating visuals first and dubbing audio afterward.
  • Efficiency at scale: MoE sparse activation lets the 114B-parameter model maintain inference costs comparable to a model with roughly 6B active parameters.
  • Extensible video: The autoregressive approach supports video continuation and interactive extension out of the box.

Who It's For

Sand.ai is aimed at creators and marketing teams that need to produce advertising spots, music videos, and short-form video at volume; researchers studying joint audio-video generation and autoregressive video models, who can also draw on the open-source MAGI-1 lineage; and developers who need video generation capabilities integrated through an API, with documentation available at platform.sand.ai.

FAQ

What is MAGI-2 Preview?MAGI-2 Preview is Sand.ai's unified 114B-parameter audio-video generation model. It generates visuals and sound jointly from a single model and, through its MagiMoE architecture, activates only 6B parameters per token.

How does joint audio-video generation differ from generating video and adding sound separately?Because one model produces both picture and sound, dialogue lines up with lip movements, actions align with sound effects, and environmental changes are matched by audio feedback. This avoids the synchronization problems common in pipelines that dub audio onto pre-generated visuals.

What camera movements can MAGI-2 produce?The model understands framing, composition, perspective, and camera intent, and can generate push-ins, tracking shots, and orbiting camera movements with pacing that approximates real filming.

Was Sand.ai's earlier model open source?Yes. MAGI-1, the predecessor to MAGI-2, was released as an open-source autoregressive video generation model, and Sand.ai has continued that open-source tradition.

Can developers integrate Sand.ai's models?Yes. Sand.ai operates a developer platform at platform.sand.ai, which provides API documentation for integrating video generation capabilities.