Sand.ai Open-Sources 114B Video Model: 10 Seconds of 1080p for Just RMB 0.50

Sand.ai has just open-sourced an MoE video generation model with 114B total parameters, while activating only about 6B parameters per inference run. Its headline selling point is straightforward: generating a 10-second, 1080p video costs approximately RMB 0.50 in inference.

If this figure can be reproduced using commonly available cloud GPUs, public weights, and standard inference configurations, its significance goes beyond being “yet another hundred-billion-parameter model.” More importantly, video generation is, for the first time, approaching a cost structure comparable to ordinary internet content rather than cinematic visual effects.

However, 114B total parameters, 6B active parameters, and RMB 0.50 are three figures that need to be considered separately. Parameter count determines model capacity, active parameters affect per-step compute, while final cost still depends on video duration, frame rate, sampling steps, parallelization strategy, hardware utilization, and whether upscaling is included. Treating these three figures as directly equivalent can easily lead to an overestimation of the benefits delivered by MoE.

114B Is Huge, but the Entire Model Does Not Need to Be Computed Every Time

The most noteworthy aspect of this model is its MoE, or mixture-of-experts, architecture.

A traditional dense model can be thought of as a company with a single enormous department: whether the task is generating people, animals, buildings, or complex camera movements, all parameters participate in the computation. MoE is more like a company with many specialized teams, where a router assigns each token or video feature to only a small subset of experts.

Sand.ai provides the following figures:

• Total parameters: 114B;
• Active parameters per inference run: approximately 6B;
• Activation ratio: approximately 5.3%;
• Target output: 10-second, 1080p video;
• Claimed inference cost: approximately RMB 0.50 per clip.

In other words, the model has knowledge and expressive capacity at the hundred-billion-parameter scale, but invokes only a small portion of its parameters at each step. It is attempting to achieve two goals that were previously difficult to combine: greater model capacity and compute costs approaching those of a small model.

However, “only 6B active parameters” does not mean “deployment is as easy as deploying a 6B model.”

The full 114B weights still need to be loaded or distributed across multiple GPUs and nodes. Even with FP8, INT8, or more aggressive quantization, weight memory, inter-GPU communication, and expert scheduling remain practical challenges. This is particularly true for video generation, where the input and output are not short text sequences but high-dimensional latents with a temporal dimension. The model must also handle cross-frame consistency, motion trajectories, and shot continuity, while attention layers, caches, and intermediate activations likewise consume substantial VRAM.

MoE therefore primarily reduces FLOPs. It does not necessarily reduce the following by the same proportion:

1. Weight storage;
2. Model loading time;
3. Cross-GPU communication;
4. Service cold starts;
5. Idle costs in low-concurrency scenarios.

MoE is highly attractive for large-scale cloud inference. For individual developers with only one or two consumer GPUs, however, a 114B model is still not something they can simply “download and double-click to run.”

What Does RMB 0.50 Mean?

Based on Sand.ai’s figures, a 10-second video costs approximately RMB 0.50, meaning the raw inference cost can be roughly extrapolated as follows:

| Generation Scale | Theoretical Inference Cost | | --- | ---: | | 10 seconds | Approx. RMB 0.50 | | 1 minute | Approx. RMB 3 | | 10 minutes | Approx. RMB 30 | | 1 hour | Approx. RMB 180 | | 1,000 ten-second videos | Approx. RMB 500 |

This is already within an acceptable range for advertising assets, e-commerce short videos, game cinematics, and social-feed content. In the past, the main problem with video models was not simply that “the quality was not good enough.” A single generation often took several minutes, and every iteration consumed costly GPU time. In commercial production, a shot that is ultimately used may go through repeated prompt revisions, random seed selection, and partial regeneration.

If producing one usable video requires an average of eight generations, the nominal RMB 0.50 becomes RMB 4 in actual asset cost. Add upscaling, frame interpolation, voice-over, lip-syncing, moderation, and storage, and the final bill continues to rise.

Therefore, “RMB 0.50 for 10 seconds” is better understood as the marginal inference cost under ideal conditions, rather than the final API retail price paid by developers. The latter must also cover idle GPU capacity, scheduling infrastructure, bandwidth, operations, maintenance, and profit.

In addition, evaluating this figure currently requires clarification of several key assumptions:

• Whether 1080p is generated natively by the model or produced by upscaling a lower-resolution output;
• The output frame rate and whether frame interpolation is included;
• Which GPU is used, along with its purchase or rental cost;
• The number of sampling steps and the numerical precision used;
• Whether the cost assumes batch concurrency and high utilization;
• Whether prompt encoding, content moderation, and retries after failed generations are included.

These variables can cause the final cost to differ by several times. At this stage, RMB 0.50 is a highly striking result, but it still needs to be reproduced by the community using the public weights.

Why Video Models Need MoE More Than Language Models Do

Language models process one-dimensional token sequences, while video models must handle both spatial and temporal dimensions. Doubling image resolution does not mean compute increases by only twofold. Longer videos also require continuous tracking of character identity, object positions, lighting changes, and camera movement.

Within a ten-second shot, a model may need to solve all of the following simultaneously:

• The subject’s appearance must not drift between adjacent frames;
• High-frequency details such as hands, faces, and text must not break down;
• Object motion must follow basic physical laws;
• Camera movement must remain consistent with scene perspective;
• If a subject is occluded by a foreground object, its identity must remain consistent when it reappears;
• The timing of actions must match the prompt description.

These capabilities cannot be obtained simply by increasing the number of sampling steps. Greater parameter capacity may indeed accommodate more motion patterns, visual concepts, and cinematic language. The problem is that if a 114B dense model is used for every inference run, it becomes extremely difficult to bring costs down.

MoE offers a more practical path: different experts can specialize in people, animals, natural environments, complex motion, or particular visual styles, and then be dynamically combined through a routing mechanism. In theory, this is more efficient than asking a single dense network to handle all visual patterns equally.

Of course, video MoE is also more difficult to tune than text MoE. If adjacent frames are routed to substantially different experts, textures, colors, and subject features may flicker. Uneven expert workloads can also leave some GPUs waiting, eroding the gains from sparse computation. Routing stability and expert-parallel efficiency will ultimately be reflected directly in video consistency and real-world throughput.

Open Source Matters More Than the “Hundred-Billion-Parameter” Label

The video generation field has no shortage of impressive demos. What it lacks are models that developers can fully inspect, deploy, and modify.

Closed-source products generally provide only a web interface or API. Developers can enter prompts, but it is difficult to know whether the model has implicitly rewritten them, whether additional upscaling was applied, or how much compute was actually used for each generation. They are also unable to continue training the model for vertical industries.

Open weights give developers at least several new possibilities:

1. Private deployment: Film, advertising, and game teams can keep unreleased assets within their own infrastructure;
2. Vertical fine-tuning: Experts can be trained around specific characters, products, cinematic styles, or industrial scenarios;
3. Inference optimization: The community can experiment with quantization, distillation, caching, and expert-parallel strategies;
4. Cost reproduction: Developers can verify whether the approximately RMB 0.50 cost for 10 seconds of 1080p video depends on specialized hardware or internal clusters;
5. Workflow integration: The model can be integrated into node-based creative tools, editing software, and automated content pipelines.

It is also important to emphasize that publishing source code, publishing weights, and permitting genuine commercial use are three different things. Before integrating the model into production, developers should carefully verify whether the license permits commercial use, where responsibility for generated content lies, the extent to which training data has been disclosed, and the distribution requirements for derivative models.

If only the inference code is released without the complete weights, or if the weights are subject to a restrictive license, the commercial value of the claimed open-source release will be greatly diminished. Conversely, if the weights, inference code, and training methods are all sufficiently complete, this release may genuinely change the cost curve for video models.

Will It Directly Replace Existing Video Generation Products?

Not in the short term.

A video model cannot be judged based on one or two sample clips. Real production use requires continuous testing of character consistency, prompt adherence, complex actions, text generation, multi-shot storytelling, first- and last-frame control, and localized editing. A model may look stunning when generating natural landscapes but quickly reveal its limitations in multi-person interactions, object collisions, or product geometry.

The advantages of closed-source products also extend beyond the foundation model itself. They typically package prompt rewriting, content safety, automatic upscaling, shot extension, audio generation, and asset management into a complete product. Even if an open-source model offers lower per-generation inference costs, development teams must still build the serving infrastructure and workflows themselves.

Sand.ai’s release is more likely to redefine the lower bound of competition: if a 114B MoE model can consistently generate 1080p video while activating only 6B parameters, the industry can no longer cite “the model is too large” as the sole justification for high video generation costs. Competition will now shift toward routing efficiency, concurrent throughput, controllable generation, and engineering integration.

Four Metrics Developers Should Actually Watch

Once the model weights and full technical details become more widely available, the following four areas will be more valuable to test than the promotional videos:

1. Cost per Usable Video

Do not calculate only the cost of a single successful generation. Instead, measure how many retries are needed on average to obtain one deliverable asset. The effective cost is usually significantly higher than the cost of a single inference run.

2. Time to First Frame and Sustained Throughput

Offline batch generation depends on throughput, while interactive products must also consider how long users wait for the first frame. If MoE requires frequent cross-GPU communication, theoretical FLOPs may be low while real-world latency remains unsatisfactory.

3. Long-Term Consistency

Ten seconds is already long enough to expose changing faces, shifting clothing, and drifting backgrounds. Per-frame clarity cannot substitute for stability over time.

4. Deployability on Mainstream Hardware

Whether a hundred-billion-parameter model can be quantized, run on a smaller number of GPUs, and support mature expert-parallel frameworks determines whether it is genuinely “community-accessible” or effectively available only to companies with large clusters.

Assessment: Video Generation Is Moving From Flexing Technical Muscle to Doing the Math

The most valuable aspect of Sand.ai’s release is not that it pushes the video model to 114B parameters, but that it places “only 6B active parameters per inference run” and “approximately RMB 0.50 for a ten-second video” on the same table.

Over the past year, video model releases have generally emphasized resolution, duration, and sample quality. The fact that cost is now front and center indicates that the industry is moving from model demonstrations toward production at scale. For developers, whether a model can reliably generate 100,000 assets per day is often more important than whether it can occasionally produce a cinematic-looking demo clip.

Our assessment is that 114B is the attention-grabbing figure, 6B active parameters are the technical core, and whether the RMB 0.50 cost can be reproduced by the community will determine the true significance of this open-source release.

If the cost figure holds, Sand.ai will push open-source video generation into a new price range. If it depends on extremely large batches, specific clusters, or excludes upscaling and retries, this is still a valuable engineering achievement—but it remains some distance from “RMB 0.50 for ten seconds, ready to use out of the box.”

As of August 5, 2026, the model has only just been released, and comprehensive third-party testing and cross-hardware reproduction remain limited. For now, the most prudent position is that it is worth downloading and testing, but production budgets should not be based solely on official sample videos and theoretical costs.

References

• SandAI-org / MAGI-1 (GitHub): The entry point for Sand.ai’s open-source video generation project, where developers can verify the code structure, deployment dependencies, license, and subsequent version updates.
• SandAI-org (Hugging Face): Sand.ai’s model community page, where developers can view public weights, model cards, and related files. Availability of specific new versions is subject to page updates.