MAGI-2 Review: The First Hundred-Billion-Parameter Open-Source Video Model With Built-In Sound

MAGI-2 is Sand.ai's open-source video generation model, released on August 5, 2026. It takes a text prompt — or a text prompt plus a starting image — and produces a 10-second 1080p video with audio that is synchronized to the action. With 114 billion total parameters in a mixture-of-experts (MoE) architecture that activates only about 6 billion per token, it is the first video model at the hundred-billion scale to publish both its weights and its inference code under the Apache 2.0 license. If you want video-with-sound generation that you can run, fine-tune, and ship inside your own stack without per-clip API billing, this is the release to watch.

What Makes MAGI-2 Different

Most open video models generate silent clips. You run the video model, then pipe the result into a separate audio model, then align the two in post. MAGI-2 removes that pipeline. Text, video, and audio tokens are concatenated into a single sequence and processed by one Transformer backbone, so speech, lip movement, body motion, ambient sound, and camera rhythm exchange information throughout the model rather than only at a narrow cross-attention step. The model renders the visuals and the soundtrack in a single pass, then muxes the audio into the output MP4.

The practical payoff is coherence: a character who starts speaking has mouth movement that matches the generated voice, and environmental sound (footsteps, water, wind) tracks the on-screen motion instead of being bolted on afterward.

Architecture: 114B Capacity, 6B Active

MAGI-2 uses a fine-grained MoE design. Each token's representation is split into 12 latent heads of 256 dimensions; each head runs its own router and selects its top-6 experts from a pool of 256 narrow experts. That yields 3,072 head-local expert units per sparse layer, of which only 72 activate per token. The result is a model with the capacity of 114B parameters but the per-token compute of roughly 6B — which is why its reported inference cost at scale is around ¥0.5 (about $0.07) per 10-second clip, roughly a tenth of comparable proprietary APIs.

Generation is two-stage: a preview stage renders at 512×896, then a refiner stage upscales to 1088×1920 (approximately 1080p). The only supported duration in this preview release is 10 seconds.

How to Run MAGI-2

MAGI-2 is self-hosted. There is no hosted API yet (one is expected), so you run it yourself:

• Hardware: 8× NVIDIA Hopper GPUs (H100 or equivalent). This is a hard requirement for the full model; quantized or distilled variants for smaller rigs are not available in the preview.
• Get the code: clone github.com/SandAI-org/MAGI-2-preview.
• Download weights: hf download sand-ai/MAGI-2-preview --local-dir ckpt (about 307 GB: 228 GB preview transformer, 14 GB refiner, 56 GB Qwen3.5-27B text encoder, plus video/audio VAEs and a distilled turbo VAE decoder). A Docker image is provided.
• Run inference: bash scripts/run_demo.sh launches the pipeline under torchrun across all visible GPUs, generating at 1080p from a samples JSON file. Output lands as sample_000.mp4, sample_001.mp4, and so on.
• Prompt enhancement (optional): MAGI-2 was trained on long, structured captions. A short prompt underuses it. The repo ships a prompt-enhancement (PE) step that calls an OpenAI-compatible LLM endpoint to expand your prompt into a structured JSON caption; leave the API key empty to skip PE and use your raw prompt.

There are no ComfyUI custom nodes for MAGI-2 at the time of writing, so expect a command-line workflow rather than a node graph.

MAGI-2 vs Other AI Video Tools

The honest takeaway: Kling AI and Runway win on accessibility and turnaround today, and Sora leads on maximum clip length. MAGI-2 wins on two things that matter for builders — it is fully open (you own the stack) and it generates synchronized audio natively. On the Artificial Analysis video leaderboard it sits around 6th, competitive with far larger proprietary systems.

Who Should Use MAGI-2

• Studios and agencies that already run GPU clusters and want to stop paying per-clip API fees for ad, short-drama, or animation production.
• Researchers who need a transparent, modifiable video foundation model to study scaling, audio-visual alignment, or fine-tuning.
• Product teams building proprietary video features who cannot send source footage to a third-party API.

If you just want to make a quick social clip, Pika, PixVerse, Seedance, or Vidu AI are faster to start. MAGI-2 is a build-it-yourself foundation, not a SaaS.

Limitations

• Hardware barrier: 8× Hopper GPUs and ~307 GB of weights make it inaccessible without serious infrastructure.
• Preview maturity: this is a preview release; the full distilled version (far fewer denoising steps) is "coming soon."
• 10-second ceiling: longer narratives require stitching multiple clips.
• No hosted API yet: you must operate the inference pipeline yourself.

FAQ

Is MAGI-2 free?

Yes. The weights and code are released under Apache 2.0, so you can use, modify, and commercialize output without per-clip licensing fees. The cost is your own GPU compute.

Who made MAGI-2?

MAGI-2 was developed by Sand.ai, following MAGI-1, its autoregressive video model from April 2025. The preview release validates that its architecture and training system scale together at the 100B level.

How much does it cost to run?

There is no API bill. Reported inference cost at scale is about ¥0.5 (≈$0.07) per 10-second clip on owned hardware, roughly a tenth of comparable proprietary APIs — but you must supply the GPUs.

Can MAGI-2 generate audio?

Yes. Audio is generated natively in the same pass as the video, with lip movement and ambient sound synchronized to the visuals.

Where do I get MAGI-2?

Weights: huggingface.co/sand-ai/MAGI-2-preview. Code: github.com/SandAI-org/MAGI-2-preview. Announcement: sand.ai/blog/magi-2-preview.

Final Assessment

MAGI-2 is the most important open video release of 2026 for one reason: it is the first hundred-billion-parameter video model you can actually own. The hardware requirement is steep, and the hosted API is not here yet, but for teams with GPUs it collapses "generate video" and "generate sound" into a single open pipeline. Watch the distilled release — that is when MAGI-2 becomes practical for most builders.