News

Sand.ai Raises $200 Million Ahead of the Full MAGI-2 Release: What the Open Preview Already Shows, and the Hardware It Needs

6 November 2026 · RenderBob team

Sand.ai said on 5 November that it raised $200 million and will release the full MAGI-2 soon. TechNode is the source. An August preview needs eight Hopper GPUs.

Chinese video-model company Sand.ai said on 5 November that it raised $200 million and will release the full MAGI-2 soon. A preview of the model was already released in August. The funding figures below are TechNode's 6 November relay of a Chinese-language Jiemian report. We have not read Jiemian or a Sand.ai post.

The funding

TechNode says Sand.ai announced a $200 million Series A+ round on 5 November, led jointly by Cornerstone Capital and Wangsu Science & Technology. The company called it unicorn status. TechNode says the company has raised $300 million this year across three rounds. No valuation is given. Proceeds are for research, training and inference infrastructure, product development and international expansion. The full model is described only as coming soon, with no date. A 22 June report put more than $100 million into two earlier rounds that year, and said the company planned to release an open-source model in July. TechNode's $300 million year-to-date figure would include that earlier money only if the rounds are the same set. We have not confirmed that from Sand.ai.

What the preview already is

ComfyUI Wiki, a community site, and 36Kr, on 5 and 6 August, describe MAGI-2 Preview. Both link to primary materials we could not open.

• Size. About 114 billion parameters in a mixture-of-experts design, with roughly 6 billion active per token.
• Output. Ten-second videos with synchronised audio in one pass, from a text prompt or a prompt plus a still image. A first stage runs at 512×896 and a refiner scales to 1088×1920.
• Availability. Weights, code and a technical report are described as public. Neither source names a licence, so terms of use are unknown.
• Claims. 36Kr says it ranks sixth on Artificial Analysis's image-to-video leaderboard, and gives Sand.ai's own estimate of about 0.5 yuan to generate 10 seconds of 1080p. Treat that as an internal cost estimate.

What it needs to run

ComfyUI Wiki says the preview needs eight NVIDIA Hopper GPUs and a checkpoint set of about 307 GB, with a distilled decoder as the default, and that no ComfyUI nodes exist for it yet. Eight Hopper GPUs and a checkpoint of that size is data-centre hardware. The earlier MAGI-1's 4.5-billion-parameter model ran on a single 24 GB graphics card, according to its Hugging Face page, so Sand.ai's releases span very different machines.

What is not known

We do not know the preview's licence, what the full MAGI-2 adds, the release date, or the quality compared with other models. The funding figure and the promise of a release soon are a single relay of Jiemian. The leaderboard rank is one data point. 36Kr's article is largely opinion and framing.

What to check before you rely on an open model

• Read the licence on the model's own repository.
• Check the hardware. Parameter counts do not say whether it fits your machine.
• Price a cloud run if you cannot run it locally.
• Test your own shots against a model you already use.
• Note the version. A preview can change.
• Keep a record of the weights version and settings for each clip.

What to watch

Sand.ai's own announcement, the full MAGI-2 release and licence, whether smaller or distilled variants appear, and independent quality and cost comparisons.

RenderBob 3 and this story

RenderBob 3 lets you choose local Apple Silicon or cloud generation per clip, so a model that only runs on data-centre hardware is a cloud choice for one shot while the rest of the cut stays local.