← Back to blog
NewsAbout 7 min read

Sand.ai Raised $200 Million. Its Video Base Model Is the Actual Product.

Published Nov 7, 2026
Sand.ai Raised $200 Million. Its Video Base Model Is the Actual Product.

On November 5, Sand.ai, the Chinese video foundation model company, announced a $200 million Series A+ round led by CAS Capital and Wangsu Data Technology, with China Time Capital, Guofang Capital, Green Pine Capital and others participating. It is the company's third round this year, bringing the ten-month total to $300 million, and it puts Sand.ai into unicorn territory officially.

The money is the headline. The more interesting part is what the company has been building with the previous two rounds: MAGI-2 Preview, an open-weight video model that treats audio, video and text as one problem instead of three.

One Transformer for Three Modalities

MAGI-2 Preview is described as a unified audio-video generation model. Text, video and audio all enter the same Transformer, and information keeps exchanging through self-attention at every layer. The architecture splits a 3,072-dimensional hidden state into 12 heads of 256 dimensions, runs them through 36 layers of a multi-head mixture of experts, and gives each head 256 experts with six selected per head. One token ends up activating 72 small experts in total.

Shared experts handle what modalities have in common. Modality-specific experts handle what they do not. Around that sits MagiMoE, a custom MoE kernel library, and MagiMuon, a distributed optimizer built for the training and inference path.

Sand.ai has released it as one of the first video base models of its size under open weights, and the pricing has been aggressive: roughly 0.5 yuan for a ten-second 1080p clip. The company also shipped MAGI-1, an autoregressive video generation model, and MagiHuman, both open source. Its consumer product, VidMuse, reportedly crossed $10 million in annualized recurring revenue two months after launch.

That sequence matters. MAGI-1 was autoregressive, MAGI-2 Preview moved to a mixture-of-experts diffusion design, and the next release, MAGI-2 final, is described as a larger version of that MoE architecture. This is a lab working through a scaling path and publishing the checkpoints as it goes, which is a different posture from companies that keep weights locked and sell API calls.

Why the video market is worth this kind of money

Look at where the demand actually sits. Kuaishou reported Kling AI revenue of RMB 850 million in the second quarter of 2026, more than 200 percent year over year, with global users past 100 million. ByteDance's Seedance 2.5 leads 13 of 20 categories in Artificial Analysis's video arena, and OpenAI has already shut down Sora, both the app and the API. Google moved developers off the Veo 3.1 Preview endpoint in October and pointed them at Gemini Omni Flash.

Meanwhile pricing is collapsing. Alibaba runs about a 30 percent promotion on Wan 3.0, which puts 1080p at roughly $0.14 per second. On OpenRouter, ByteDance holds 26.1 percent of video requests, HeyGen 19.7 percent, Google 18.5 percent. TBC's mischief 1, a Wan 2.2 derivative pitched at half a cent per second, jumped 481 percent week over week the moment it launched.

When unit prices fall toward zero, the money moves to whoever owns the base model. A company that can train and serve video generation at the margin captures the volume that everyone else is renting from Google, ByteDance and Alibaba.

Why unified audio-video matters commercially

One transformer for audio and video removes production steps, and that is the reason to build one.

The standard pipeline for a product spot today goes like this: generate the picture, generate or record the voice, cut music, mix, and reconcile lips to dialogue. Every one of those is a separate tool with its own cost and its own failure mode, and the reconciliation step is where projects stall. When a model produces dialogue, ambience and action sound in the same pass as the image, the lip sync is a property of the generation rather than an edit performed afterward.

Every vendor in this category now claims that capability, and it decides whether AI video can carry a talking-head product explainer or a branded story spot. It is also the capability where a unified architecture has the clearest theoretical advantage, since the model does not have to describe sound in text and hope the picture matches.

Sand.ai's MAGI-1 was autoregressive. MAGI-2 Preview moved to a diffusion-based mixture of experts. Both were released, which means anyone can look at how the generation quality changes between a sequential model and a parallel one, and at what happens to audio sync when the architecture changes.

How it stacks up against the rest of the field

Compare the price points rather than the claims. Alibaba runs Wan 3.0 at about 0.3, 0.6 and 1.2 yuan per second at 480p, 720p and 1080p, with a roughly 30 percent promotion that pushes 1080p to $0.14 per second on resale routes. MiniMax open-sourced H3-Base in August with 2K output, 15-second clips and native stereo, and priced it at roughly one third of comparable flagship models. TaoMate-H3, an open-source effort from Alibaba's commerce side, pushed further and made time-to-first-segment its headline metric rather than price per second.

Every one of these is an attack on the same line item. Sand.ai's 0.5 yuan for a ten-second 1080p clip sits inside that band rather than below it, which suggests the strategy is breadth and openness rather than a price war it cannot win against a company with Alibaba's infrastructure.

The uncomfortable questions

None of this is independently verified. Sand.ai reports its own ARR and its own round size, and the PresentBench-style comparisons that matter to buyers are usually vendor-published. The MAGI-2 architecture is genuinely interesting, but a unified audio-video transformer with 72 activated small experts per token is also expensive to train, and the company is funding that with money it raised rather than with revenue it has earned.

There is a licensing angle worth watching too. Several Chinese open video models shipped in 2026 with terms that exclude the United States from commercial use, which pushes Western developers toward closed alternatives and complicates any open ecosystem story. Sand.ai's existing open releases appear to follow the more standard permissive path, which is part of why the round was oversubscribed by the investors it named.

And the commercial product is the part to watch. VidMuse reaching $10 million ARR two months after launch is a real number, but video creation platforms have a habit of growing on one-off campaigns and then flattening out once the novelty wears off. The open model is what makes the company durable; the app is what funds the research.

What to watch next

MAGI-2 final is described as imminent, with a larger MoE and the same unified audio-video design. If it holds the 0.5 yuan per ten-second 1080p price while pushing longer clips and better lip sync, the model becomes a plausible training target for anyone building video products on a budget.

Watch three things in particular. Whether the weights actually ship under a permissive license. Whether the audio track holds up when a clip has dialogue, since native audio is now table stakes in this category. And whether VidMuse's revenue keeps climbing after the launch quarter, or whether it tracks the general AI video spend curve that every platform in this market is riding.

For now, the useful takeaway is simpler than the fundraising headline. Sand.ai is building video base models the way the open community expects labs to build them: publish the checkpoints, publish the architecture details, compete on price, and sell the hosted version to fund the next one.

Related articles