MAGI-2 Preview: Open-Source 114B Video Model

MAGI-2 Preview: The Open Video Model That Activates Only 6B Parameters

Video generation has a scaling problem. Better quality usually means larger models, longer context, more denoising steps, and much higher inference bills. SandAI’s MAGI-2 Preview takes a different route: keep the model’s total capacity large, but activate only a small fraction of it for each token.

The result is a 114B-parameter unified audio-video generation model that activates approximately 6B parameters per token. SandAI calls the architecture MagiMoE, an ultra-fine-grained mixture-of-experts system designed specifically for long video sequences and multimodal generation.

The project is not a lightweight consumer app. The released inference code currently requires eight NVIDIA Hopper GPUs and roughly 307GB of downloaded checkpoint files. But it is significant for a different reason: MAGI-2 makes a serious attempt to show how open video models can scale without paying dense-model costs at every step.

What MAGI-2 Preview is

MAGI-2 is a unified audio-video model. Text, video, and audio are processed as one token sequence through a shared Transformer backbone rather than through completely separate modality towers.

That design matters because video is not just a sequence of images. A useful video model must learn relationships between:

• Objects and their movement
• Camera motion and scene geometry
• Speech, lip movement, and facial expression
• Environmental sound and visible events
• Music, rhythm, and temporal structure

A single-stream architecture gives these modalities a common representation. The goal is not merely to generate attractive frames, but to model the event as a coordinated audiovisual sequence.

MAGI-2 Preview supports both text-to-video and image-to-video generation. The current release generates 10-second clips, with audio generated alongside the video and muxed into the final output. The pipeline operates in two stages: a preview model denoises at lower resolution, and a refiner brings the result up to 1080p.

The MagiMoE architecture

A dense 114B model would activate almost all of its parameters for every token. That is expensive for any workload, but especially painful for video, where the system processes a large number of spatial and temporal tokens and repeatedly runs a denoising network.

MAGI-2 uses an ultra-fine-grained MoE approach instead. It stores approximately 114B parameters in total but activates only around 6B per token. The router selects the relevant expert capacity for each part of the sequence, allowing the model to preserve a large parameter pool without paying the full dense-model compute cost at every operation.

This does not mean MAGI-2 is equivalent to a 6B model. Total parameters still provide capacity, while active parameters influence the amount of computation used for each token. Attention, routing, communication, video decoding, and the repeated denoising schedule still contribute to the final cost.

That distinction is important. “6B active” is an efficiency feature, not a guarantee that the model will run on a 6GB GPU.

Why ultra-fine-grained experts matter for video

Traditional MoE systems were largely developed around language-model workloads. Video generation adds several complications:

1. The sequences are much longer.
2. Tokens represent different spatial, temporal, audio, and text signals.
3. Denoising repeatedly revisits the representation.
4. Expert routing can create severe communication and load-balancing problems across GPUs.
5. Activations and intermediate video latents can consume more memory than the model weights alone.

SandAI’s approach co-designs the architecture with the training and inference systems. The model, routing strategy, kernels, communication patterns, and data pipeline are treated as one scaling problem rather than separate optimizations added later.

The preview release should therefore be read as a systems research milestone, not just as a parameter-count announcement.

Open source means “research-accessible,” not “runs on a laptop”

SandAI has released the inference repository under Apache-2.0, published the technical report, and linked the model weights on Hugging Face. The repository includes Docker instructions and the code for text-to-video and image-to-video inference.

The current hardware requirement is substantial:

• Eight NVIDIA Hopper GPUs
• Python 3.12
• A recent CUDA toolkit
• ffmpeg for muxing audio and video
• MagiAttention and MagiCompiler dependencies

The full checkpoint download is approximately 307GB. The repository breaks that down into roughly 228GB for the preview-stage transformer, 56GB for the Qwen3.5-27B text encoder, 14GB for the refiner, and smaller VAE/audio components.

This is a much more honest picture than calling MAGI-2 a local consumer model. It is open and reproducible for teams with serious GPU infrastructure, but it is not currently a MacBook or single-GPU workflow.

Generation quality and current limitations

MAGI-2 Preview is an intermediate research release. SandAI says the release validates that its architecture, training system, and data methodology can scale to the 100B level, while also identifying areas that need further research.

There are practical limitations in the current version:

• Clips are currently limited to 10 seconds.
• The base model uses 100 preview denoising steps plus 5 refiner steps.
• The model has not yet been step-distilled, so denoising remains the main source of wall-clock latency.
• The distilled release is listed as coming later.
• The complete checkpoint footprint is too large for ordinary local hardware.

The absence of step distillation is particularly relevant. Sparse activation reduces per-token model computation, but video generation still performs many denoising iterations. A future distilled version could have a larger practical impact on generation time than simply adding more GPU memory.

Prompting MAGI-2

The repository warns that short prompts underuse the model. MAGI-2 was trained on long, structured captions that describe the scene, actions, camera, lighting, sound, and temporal progression.

The inference pipeline includes optional prompt enhancement for both text-to-video and image-to-video. An OpenAI-compatible language-model endpoint can rewrite a short request into a structured caption before it reaches the video model.

For example, instead of writing:

“a dog running on a beach”

A stronger prompt would specify the subject, movement, environment, camera, time, and sound:

“Ten-second cinematic shot of a golden retriever sprinting along a wet shoreline at sunrise. The camera tracks from a low side angle as the dog splashes through shallow waves, ears moving with each stride. Warm backlight, realistic fur motion, distant gulls, rhythmic surf, natural handheld micro-movement, continuous action with no cuts.”

This is not merely a prompt-writing preference. Structured descriptions better match the data representation the model was trained to consume.

MAGI-2 versus mainstream video APIs

SandAI’s announcement positions MAGI-2 around a much lower generation cost, with a claim of roughly one-tenth the cost of mainstream models. That figure should be treated as a company claim rather than a universal price comparison.

The real comparison depends on what is included:

• GPU rental price
• Number and type of GPUs
• Denoising steps
• Output resolution
• Audio generation
• Failed generations and retries
• Engineering and storage overhead
• API markup and service availability

An open model can reduce per-clip marginal cost when a team already owns or leases the required GPUs. It also offers control over data, deployment, fine-tuning, and output pipelines. But the initial infrastructure cost is significant, and a managed API may still be cheaper for occasional creators.

Who should care about MAGI-2?

MAGI-2 Preview is relevant to:

• Researchers studying scalable video architectures
• Video-model teams building on open weights
• Studios that need private generation pipelines
• Infrastructure engineers optimizing multi-GPU inference
• Developers experimenting with audio-video foundation models
• Teams that want to fine-tune or customize a model rather than consume an API

It is not yet the best choice for:

• A single consumer GPU
• A laptop-based creative workflow
• Instant short-form generation
• Developers who want a simple hosted API
• Projects that require arbitrary clip lengths out of the box

Final verdict

MAGI-2 Preview is interesting because it attacks the right bottleneck. Instead of making every token pay for a dense 100B-plus network, SandAI builds a large expert pool and activates only approximately 6B parameters at a time.

That does not make video generation free, and it does not make the model small. The release still requires eight Hopper GPUs, a 307GB checkpoint footprint, and a multi-stage denoising pipeline. But it demonstrates a credible path toward larger open video models with more manageable per-token computation.

The next important release will be the distilled version. If SandAI can preserve the model’s audio-video quality while reducing the 100-plus denoising steps, MAGI-2 could move from an impressive research preview toward a more practical production system.

For now, the verdict is simple: MAGI-2 Preview is not a laptop video generator. It is an open infrastructure blueprint for scaling audiovisual generation without activating the entire model every time.

Sources: SandAI technical report, MAGI-2 Preview GitHub repository, MAGI-2 Preview weights.

Visual: Model or tool execution path

This original diagram condenses the runtime path readers need to reason about.

diagram

Visual reading: model output is not automatically trusted. Tool calls, retrieved context, and generated code need a validation boundary before execution or publication.