MAGI-2 Preview is a public research release of a unified audio-video generation model developed by Sand.ai. Published on August 5, 2026, it accepts either a text prompt or a prompt plus a still image, then generates a ten-second video with an accompanying audio track. Sand.ai reports approximately 114 billion parameters in the main model, with about 6 billion parameters activated for each token through a mixture-of-experts architecture called MagiMoE.[1][2]

The word Preview is part of the official name and describes the release status. Sand.ai presents the model as an intermediate test of a scaling approach, not a finished MAGI-2 product. The public package includes inference code, configuration files, model weights, and a Docker image. It does not include training code, a training dataset, or a quantitative evaluation suite. Sand.ai's web article is the technical release description; no standalone arXiv paper or PDF technical report accompanied the release.[1][2][3]

Quick facts

Field	Details
Developer	Sand.ai
Release date	August 5, 2026 [1]
Supported inputs	Text prompt, or text prompt plus a first-frame image [2]
Generated output	Ten-second video with jointly generated audio [2]
Main model size	Approximately 114B total parameters, approximately 6B active per token [1]
Core architecture	Forty-layer single-stream Transformer with fine-grained MoE in 36 middle layers [1][4]
Reference delivery tier	1080 by 1920 portrait output after preview and refinement stages [2]
Reference hardware	Eight Hopper-generation Nvidia GPUs [2]
Public artifacts	Inference code, checkpoints, configs, Docker image, and model card [2][3]
Top-level license labels	Apache-2.0 for the Sand code and model repositories [2][3]

Architecture

MAGI-2 Preview places text, video, and audio tokens in one sequence and processes them with the same Transformer backbone. This is a single-stream design: the modalities interact through self-attention throughout the network instead of using separate towers joined by a small number of cross-attention layers. Sand.ai says this interface was carried over from daVinci-MagiHuman, its earlier synchronized audio-video model.[1][7]

The preview backbone has 40 Transformer layers and a hidden width of 3,072. The first two and last two layers are dense. The 36 layers between them use Multi-Head MoE. A learned projection divides each token representation into 12 routed subspaces, each 256 dimensions wide. Every subspace has its own router and a pool of 256 narrow experts. Each router selects six experts for that part of the token.[1][4]

This produces 3,072 head-local expert units in each sparse layer, of which 72 are selected for one token. Those units are not 3,072 conventional full-width experts. Each operates only on a 256-dimensional slice, and different slices of the same token can choose different expert combinations. The design extends earlier Multi-Head Mixture-of-Experts work, which split token representations into sub-tokens, and Multi-Head LatentMoE, which gave individual latent heads independent routers and expert pools.[8][9]

The distinction between total and active parameters matters. The approximately 114B figure describes stored model capacity, while approximately 6B is Sand.ai's estimate of the parameters used for a token. It does not mean the checkpoint occupies the space of a 6B model. It also does not equal end-to-end floating-point operations because attention, routing, projections, decoding, and supporting models still contribute computation. Sand.ai explicitly notes that sparse expert routing does not remove attention cost.[1]

MAGI-2 Preview also includes a separate refiner. Its published configuration specifies a dense 30-layer Transformer with a hidden width of 4,096 and local window attention. The 114B headline should therefore not be read as a count of every parameter in the complete inference pipeline, which also loads this refiner, a text encoder, and several encoders or decoders.[2][3][5]

Distributed execution and training design

MagiMoE pairs the head-routed architecture with a distributed execution scheme. In standard expert parallelism, routing decisions determine how many token copies travel to each device. That can create variable communication volumes, uneven memory use, and stragglers. In the Head Parallel design used here, complete head partitions are assigned across devices before local expert routing. Cross-device transfers therefore have fixed shapes that can be allocated in advance, while the data-dependent part of routing stays local to each head owner.[1][8]

Sand.ai describes a hierarchical implementation for training. InfiniBand carries fixed-shape head activations between nodes, while NVLink is used within a node to gather and reshard expert weights, gradients, and optimizer state. Router scores are computed in FP32, the main expert computation uses BF16, and routing, token reordering, expert feed-forward computation, and output merging are fused to reduce overhead. Load balancing uses an auxiliary-loss-free expert bias rather than adding a separate balancing loss to the main objective.[1]

The company also describes MagiMuon, a hybrid optimizer designed around the head-by-expert matrix layout. It applies Muon-style orthogonalization separately to batches of expert matrices, while adapters, attention sinks, gating parameters, and several other parameter families receive Adam-style updates. These are architecture and systems disclosures, not a reproducible training recipe. The release does not state the training cluster size, accelerator count, training duration, optimization-step count, total compute, or energy use.[1]

Data methodology

Sand.ai says the project's data process moved away from aggressively filtering videos merely because they were difficult for a smaller generator to model. Its release report says the pipeline retained complex motion, interactions among multiple people, shot changes, subtitles, on-screen text, dialogue, environmental sound, and music where possible, then applied more detailed multimodal annotation. It also says safety, compliance, privacy, corruption handling, and deduplication remained part of data governance.[1]

Those statements describe a strategy rather than an auditable training dataset. Sand.ai did not publish dataset names, source domains, sample counts, hours of footage, language distribution, consent methods, copyright review results, or a license breakdown. It also did not release a dataset manifest. Claims in the report about better identity continuity across shots and the appearance of dialogue or subtitle effects are expressly described as qualitative observations that still require controlled study.[1][2]

Generation pipeline

The released pipeline supports text-to-video and image-to-video generation. In the second mode, the image is treated as the first-frame reference. Video and audio latents are generated together, decoded, and written to an MP4. The pipeline invokes ffmpeg to mux the audio track; if ffmpeg is unavailable, it still writes a silent video.[2]

Generation has two denoising stages. The base magi2_preview model runs 100 steps at low resolution. For the reference 1080p portrait preset, this stage works at 512 by 896. The magi2_refiner stage then runs five steps and produces a 1088 by 1920 internal frame size. The output path can rescale that result to exact 1080 by 1920. The entry point also exposes 272p and 540p presets.[2][4][5]

Only ten-second clips are supported by the released model. Neither the preview transformer nor the refiner is step-distilled, and the repository labels a lower-step distilled release as coming later. The published base configuration therefore uses 105 model steps in total. Sand.ai does not provide an official latency or throughput measurement for this configuration.[2]

Prompt enhancement is optional. When configured, the reference client calls an OpenAI-compatible language-model endpoint and asks it to turn a short request into a structured description of a ten-second clip. Without an API key, the pipeline sends the user's original prompt directly to its text encoder.[2]

Released artifacts and hardware

The Hugging Face repository is about 307 GB. Its largest component is the 228 GB preview transformer, distributed across 56 Safetensors shards. The package also contains a roughly 13.5 GB refiner, a 56 GB Qwen3.5-27B text encoder, a Wan2.2 video VAE, a Stable Audio Open 1.0 audio VAE, and a turbo VAE decoder. These supporting components explain why the download size and deployment requirements cannot be inferred from the 6B active-parameter figure.[2][3]

The reference setup requires eight NVIDIA Hopper GPUs, Python 3.12, a recent CUDA toolkit, and ffmpeg. The configuration uses eight-way context parallelism and eight-way expert parallelism. Components can be staged between CPU and GPU across phases, and the preview and refiner default to a round-trip offload mode. Sand.ai says this is necessary because, at the 1080p tier, the two models cannot both remain beside the other's activations on a single 80 GB card.[2][4]

The repository does not specify a particular Hopper model, a minimum aggregate VRAM figure, or a supported consumer-GPU configuration. Community reports about alternative hardware are not part of the official compatibility statement and should not be treated as established requirements.

Evaluation and evidence limits

The release report does not include a MAGI-2 benchmark table, a human preference evaluation, baseline comparisons, a scaling curve, an ablation study, a throughput benchmark, or a safety evaluation. Sand.ai says more complete scaling curves, controlled ablations, and analyses of capability boundaries are future work.[1]

This evidence gap prevents factual claims that MAGI-2 Preview has better visual quality, prompt adherence, physical consistency, or audio synchronization than another model. The release page shows and discusses qualitative behavior, but it labels the observations as preliminary. Parameter count and sparse activation are verified architectural properties, not measures of output quality. Descriptions such as top-tier, world-first, or fully open source should therefore be understood as promotional positioning unless separately demonstrated.

Relationship to MAGI-1

MAGI-1, released in 2025, generates video by autoregressively predicting fixed-length temporal chunks. Its largest documented variant has 24B parameters, and its chunked design supports video continuation and streaming generation.[6] MAGI-2 Preview belongs to the same Sand.ai research program, but it is not a straightforward enlargement of that architecture.

Sand.ai explicitly says it did not carry MAGI-1's autoregressive chunking into MAGI-2 Preview. Combining chunked autoregression, unified audio-video modeling, and 100B-scale distributed training would have introduced too many variables for the scaling experiment. Instead, MAGI-2 starts from daVinci-MagiHuman's simpler single-stream audio-video interface, then adds the fine-grained MoE capacity and associated systems work.[1][7]

Licensing, safety, and limitations

Sand.ai labels both its code repository and model repository Apache-2.0.[2][3] That top-level metadata is not necessarily a complete license analysis of every bundled dependency. The package identifies its audio VAE as Stable Audio Open 1.0, while Stability AI labels the upstream Stable Audio Open 1.0 model with the Stable Audio Community License. Sand.ai's release documentation does not explain how those terms interact with its top-level Apache label, so downstream users should review component licenses for their intended use.[2][10]

The published model card does not include a dedicated risk taxonomy, misuse evaluation, content-filtering specification, output watermark, or provenance mechanism. The report's statement that safety and privacy checks were part of data governance supplies no methods or audit results. Together with the missing dataset disclosure, this limits independent assessment of copyright, consent, representation, and memorization risks.[1][2][3]

Practical limitations are more concrete. The released model generates only ten-second clips, requires a 307 GB checkpoint bundle, uses 105 denoising and refinement steps, and targets an eight-Hopper setup. Its output quality has not been established through published quantitative or independent evaluation. MAGI-2 Preview is therefore best understood as an open-weight architecture and inference research release, not a validated general-purpose production system.

References

Sand.ai. MAGI-2 Preview: Scaling Video Generation Models Efficiently. August 5, 2026. https://sand.ai/blog/magi-2-preview
SandAI-org. MAGI-2-preview GitHub repository and README. Accessed August 14, 2026. https://github.com/SandAI-org/MAGI-2-preview
Sand AI. sand-ai/MAGI-2-preview model repository and model card. Hugging Face. Accessed August 14, 2026. https://huggingface.co/sand-ai/MAGI-2-preview
SandAI-org. magi2_preview.json. MAGI-2-preview configuration. Accessed August 14, 2026. https://github.com/SandAI-org/MAGI-2-preview/blob/main/configs/magi2_preview.json
SandAI-org. magi2_refiner.json. MAGI-2-preview configuration. Accessed August 14, 2026. https://github.com/SandAI-org/MAGI-2-preview/blob/main/configs/magi2_refiner.json
Sand.ai et al. MAGI-1: Autoregressive Video Generation at Scale. arXiv:2505.13211, May 19, 2025. https://arxiv.org/abs/2505.13211
SII-GAIR and Sand.ai. Speed by Simplicity: A Single-Stream Architecture for Fast Audio-Video Generative Foundation Model. arXiv:2603.21986, March 23, 2026. https://arxiv.org/abs/2603.21986
Chenwei Cui et al. Multi-Head LatentMoE and Head Parallel: Communication-Efficient and Deterministic MoE Parallelism. arXiv:2602.04870, February 4, 2026. https://arxiv.org/abs/2602.04870
Xun Wu et al. Multi-Head Mixture-of-Experts. arXiv:2404.15045, April 23, 2024. https://arxiv.org/abs/2404.15045
Stability AI. Stable Audio Open 1.0 model card and license metadata. Accessed August 14, 2026. https://huggingface.co/stabilityai/stable-audio-open-1.0

Editing: MAGI-2 Preview