StepVideo-T2V

What is StepVideo-T2V?

StepVideo-T2V is an open-weight text-to-video generation model released by StepFun on February 17, 2025. Its primary job is to create video from a written prompt, such as a description of a scene, action, camera movement, or visual style. The model supports prompts in both English and Chinese, giving it a broader language range than a system designed around only one prompt language.

The model has 30 billion parameters and is intended for self-hosted inference rather than a hosted, pay-per-use API. StepFun provides the model weights and inference implementation through its official repositories, and the release is distributed under the MIT license. That licensing and deployment model make StepVideo-T2V particularly relevant to research, experimentation, and organizations that want control over their own video-generation environment.

StepVideo-T2V belongs to StepFun’s Step-Video model family. The official release also includes a separate Step-Video-T2V-Turbo checkpoint, but that is a distinct model and should not be treated as part of the specifications for the standard StepVideo-T2V record discussed here.

How the model generates video

StepVideo-T2V uses several components designed specifically for video rather than treating generation as a sequence of unrelated images. A Video-VAE, or video variational autoencoder, compresses the visual data before generation and reconstructs it afterward. The documented compression is 16x16 spatially and 8x temporally. In practical terms, the model works in a more compact representation than the full-resolution video, which reduces the amount of visual data that the generative network must process.

The generative core is a diffusion Transformer, commonly abbreviated as DiT. Its architecture uses 48 layers and 48 attention heads per layer, with 3D full attention across the video representation. The 3D design is important because video has a time dimension in addition to height and width. Attention across these dimensions helps the model connect objects and actions across successive frames instead of generating every frame in isolation.

The release uses Flow Matching as part of its generation process. It also describes video-based Direct Preference Optimization, or Video-DPO. DPO is a preference-training method that adjusts a model toward outputs judged more desirable; in this case, the optimization is applied using video preferences. These design choices are provider-documented technical characteristics, not guarantees that every generated clip will have consistent motion or cinematic quality.

Text conditioning is handled with bilingual text encoders. This allows the model to interpret English and Chinese prompts without requiring users to translate them into a single supported language. The available research does not provide a separate quality comparison between the two languages, so language support should not be interpreted as an assurance of identical results for every prompt.

Documented output and hardware limits

The official repository documents generation at 544 × 992 resolution for 204 frames and reports a peak GPU-memory requirement of approximately 77.64 GB when flash attention is enabled. It recommends an 80 GB GPU. These figures are especially important for planning deployment: although the weights are openly available, running the standard model is not a lightweight local task.

The supplied research does not specify a token-style context window, a maximum prompt length, or a maximum output-token setting. It also does not establish a general resolution or frame limit beyond the documented 204-frame example. Therefore, users should consult the current repository configuration before assuming that other resolutions, longer clips, or different memory-saving settings are officially supported.

Strengths and practical benefits

• Open-weight deployment: The model can be downloaded and run with the provided inference code instead of requiring access to a proprietary hosted service.
• Large video-focused architecture: Its 30-billion-parameter DiT and 3D attention design are built around temporal visual data rather than ordinary text generation.
• Bilingual prompting: English and Chinese prompts are supported by the documented text-encoding setup.
• Longer documented clips: The release supports generation up to 204 frames in its documented configuration.
• Permissive license: The MIT license is useful for projects that need an open distribution model, subject to the license terms and any separate obligations associated with the implementation or dependencies.

For an intermediate user, the most meaningful advantage is the combination of model access and deployment control. An organization can investigate the inference process, keep prompts and generated material inside its own environment, and adapt the surrounding workflow without depending on a commercial generation endpoint. That flexibility comes at the cost of substantial hardware and operational complexity.

Limitations and trade-offs

The main limitation is the resource requirement. A peak requirement of about 77.64 GB of GPU memory places the standard model in the workstation, server, or cloud-GPU category. It is not a practical choice for most laptops, standard desktops, or low-memory accelerators. Even when the model fits, generation speed will depend on the GPU and inference configuration; the supplied research does not provide a verified frames-per-second figure or a fixed generation-time estimate.

StepVideo-T2V also lacks the convenience of a managed commercial API in the documented release. No official hosted input or output price was identified. The effective cost is therefore the cost of acquiring or renting suitable compute, storing the weights and generated videos, and maintaining the inference environment. This can be economical for sustained or private workloads, but it is less convenient than paying for occasional generations through a hosted service.

The model is text-to-video only in the documented configuration. Text is the supported input modality, and video is the direct non-text output. The research does not document image-to-video, video-to-video, audio input, native audio generation, tool use, streaming output, structured JSON output, or function calling. A generated video should not be assumed to contain synchronized speech, sound effects, or music.

There is also no supplied benchmark evidence for coding ability, reasoning performance, visual fidelity, temporal consistency, or prompt adherence. Because StepVideo-T2V is a video generator rather than a general-purpose language model, conventional coding or reasoning scores are not meaningful primary selection criteria. The model may interpret complex prompts, but the research does not justify presenting it as a reasoning or coding system.

Pricing and access

StepVideo-T2V has no official hosted API price in the supplied research. The model weights and inference code are available for self-hosted use, so there is no provider-listed per-token, per-image, or per-video generation fee to report. Users must budget for compatible NVIDIA GPU hardware or cloud instances, along with storage and operational costs.

This pricing model creates a clear trade-off. For occasional experimentation, renting an 80 GB-class GPU may be less convenient or more expensive than using a hosted video service. For sustained internal generation, privacy-sensitive workflows, or research requiring repeated access, self-hosting may provide more control and potentially more predictable infrastructure economics. Those conclusions are deployment considerations rather than published StepFun pricing claims.

When to choose StepVideo-T2V

Choose StepVideo-T2V when open weights, local control, and model experimentation matter more than turnkey access. It is a reasonable candidate for:

• research into open text-to-video architectures and video diffusion systems;
• self-hosted creative video production from English or Chinese prompts;
• private or controlled environments where prompts and outputs should remain on organization-managed infrastructure;
• developers who want to inspect or adapt an MIT-licensed video-generation release; and
• experiments involving longer clips within the documented 204-frame configuration.

Another option may be more appropriate when the priority is fast, low-maintenance generation, access from modest hardware, or a predictable hosted price. A managed video API is generally easier for occasional users and production teams that do not want to operate large GPUs. A lighter or accelerated video checkpoint may be preferable when throughput is more important than using the standard 30-billion-parameter model, although the supplied research does not provide enough evidence to compare the separate Turbo checkpoint quantitatively.

StepVideo-T2V is also a poor fit for workflows that require native audio, image-to-video editing, video continuation, real-time response, tool calling, or structured machine-readable output. Those capabilities are not documented for this model. Users should select a system that explicitly supports the required input and output modes rather than assuming that an open text-to-video model provides them automatically.

Bottom line

StepVideo-T2V is a substantial open-weight text-to-video release from StepFun, with a 30-billion-parameter video architecture, bilingual text prompting, and documented generation up to 204 frames. Its Video-VAE, 3D-attention diffusion Transformer, Flow Matching, and Video-DPO components make it technically interesting for video-model research and self-hosted experimentation.

Its practical cost is hardware: the documented configuration requires roughly 77.64 GB of GPU memory and recommends an 80 GB NVIDIA GPU. There is no official hosted API price, and the model does not come with documented audio, image, tool, streaming, or structured-output features. In short, StepVideo-T2V is best understood as an open research and deployment asset for users prepared to manage large-scale inference, not as a lightweight consumer video tool or a general-purpose multimodal assistant.