NextStep-1.1

What is StepFun NextStep-1.1?

StepFun NextStep-1.1 is an open-weight text-to-image model developed by StepFun’s Multimodal Intelligence team. Its primary job is to turn a written prompt into an image. The published example uses a 512-by-512 generation configuration and exposes controls such as sampling steps, guidance values, random seed, image count, positive prompts, and negative prompts.

The checkpoint is intended for local or self-hosted inference. Its weights and related materials are distributed through Hugging Face and ModelScope, while the official NextStep-1 repository provides the inference pipeline and setup guidance. This makes the model a different proposition from a hosted image API: users obtain and operate the model themselves instead of sending image requests to a metered, first-party endpoint.

Within StepFun’s catalog, NextStep-1.1 is the post-trained release in the NextStep-1.1 family. It should not be confused with the separately listed NextStep-1.1-Pretrain variants or older NextStep-1 checkpoints. StepFun describes the release as a re-engineered model with extended training and flow-based reinforcement-learning post-training, rather than simply a conventional fine-tune.

How the architecture works

Many image generators use diffusion-style denoising, where an image is gradually refined from noise. NextStep takes an autoregressive approach: it generates a sequence that combines text tokens with continuous image tokens. A causal Transformer processes this sequence, while a lightweight flow-matching head predicts the continuous visual tokens used to form the image.

“Continuous image tokens” means the visual representation is not limited to a sequence of discrete codebook entries. In practical terms, the architecture is designed to preserve richer numerical information while the model predicts image content. Readers do not need to reproduce the architecture to use the checkpoint, but this design is important for understanding why NextStep-1.1 is positioned as an autoregressive image-generation system rather than a standard diffusion-only model.

StepFun says version 1.1 adds extended training and flow-based reinforcement learning. The stated goals are higher visual fidelity, improved image texture, fewer visual artifacts, and better numerical stability during reinforcement-learning training. These are provider claims about the training approach; the supplied research does not provide an independent benchmark table that quantifies the improvement over NextStep-1.

There is a small naming distinction in the published specifications. The Hugging Face model card lists approximately 15 billion parameters, while the repository’s broader architecture description refers to a 14-billion-parameter autoregressive backbone with a lightweight flow-matching head. These figures describe related parts of the system rather than establishing a different model identity.

Capabilities and supported modalities

The verified primary capability of this checkpoint is text-to-image generation. It accepts text prompts and produces images. The model should therefore be evaluated as a visual generation system, not as a general conversational model.

The wider NextStep project concerns research involving text and visual tokens, but that broader scope should not be used to infer image editing, multimodal understanding, video, speech, or agent features for the NextStep-1.1 checkpoint. The available model materials specifically support treating it as a prompt-driven image generator.

Quality, speed, and cost trade-offs

NextStep-1.1’s main attraction is control. Downloadable weights allow an organization to run the model in its own environment, inspect the pipeline, and adapt the surrounding software for research or production experiments. The Apache 2.0 license listed on the Hugging Face model card may also be useful for teams evaluating commercial or internal deployment, although users should review the complete model-card, repository, and dependency terms before deployment.

The trade-off is operational complexity. A model of this size requires substantial GPU memory and a compatible software environment. The published setup uses Python, PyTorch, Transformers-compatible components, CUDA-capable hardware, safetensors, and custom code. Actual generation speed depends on the GPU, implementation, sampling configuration, and output resolution. No official throughput figure is established in the supplied materials.

Self-hosting can be more economical than paying for repeated hosted image requests when a team already has suitable hardware or needs high-volume local experimentation. It is not automatically cheaper for occasional users, however. Hardware acquisition, storage, engineering time, maintenance, and electricity are part of the practical cost. Users who want a simple prompt-and-download workflow may prefer a managed image service, even if it offers less control.

The available research does not identify an official per-image, per-token, or subscription price for NextStep-1.1. It is therefore best described as having no published first-party hosted API price rather than as a free hosted service. Downloadable weights do not eliminate the cost of running the required infrastructure.

Context, output, and other limitations

Traditional language-model limits such as context-window length and maximum output tokens are not meaningful published specifications for this image-generation checkpoint. The model card and supplied research do not establish a context length or a maximum output-token count. Image dimensions, sampling steps, and generation behavior depend on the local pipeline and its configuration; the demonstrated example uses 512-by-512 output but should not be treated as a universal maximum resolution.

• Hardware: The model is large and generally requires substantial GPU resources for practical local inference.
• Software: The official workflow uses custom code and a compatible Python, PyTorch, Transformers, and CUDA environment.
• Hosted access: The model is not identified as deployed through a Hugging Face Inference Provider, and no official managed endpoint is established.
• Modalities: Published support is centered on text-to-image generation, not speech, video, embeddings, or general multimodal interaction.
• Configuration: Quality and speed vary with sampling steps, guidance settings, seed, hardware, resolution, and implementation.
• Evaluation: The supplied sources describe intended improvements but do not provide independent benchmark results that can verify a particular quality ranking.

Best use cases

NextStep-1.1 is a strong candidate for teams that specifically want to study or deploy open-weight autoregressive image generation. Suitable projects include local prompt-to-image tools, research into continuous visual tokens, reproduction of StepFun’s published approach, experimentation with flow-based reinforcement-learning methods, and creative workflows where image generation must remain inside a controlled environment.

It may also suit developers who need to modify the inference pipeline rather than consume a fixed commercial API. For example, a research group could vary sampling settings, compare checkpoint behavior, or integrate the model into an internal creative application while keeping prompts and outputs on its own infrastructure.

When to choose NextStep-1.1

Choose NextStep-1.1 when open weights, local execution, and architectural experimentation matter more than turnkey access. It is particularly relevant if your team has CUDA-capable hardware, can manage a custom inference environment, and wants to investigate an autoregressive alternative to more common diffusion-based image generators.

Choose a hosted image-generation service instead when you need predictable access without managing GPUs, model files, dependencies, or updates. A hosted option may also be more appropriate for occasional usage where self-hosting costs more than paying per generated image.

Choose a conversational or multimodal model instead when the main task involves dialogue, document analysis, image understanding, speech, tool use, or structured text generation. NextStep-1.1 is not documented as a general assistant and does not provide published support for those functions.

Finally, choose a different local image model if your priority is a documented editing workflow, image-to-image input, a smaller hardware footprint, or published performance measurements that are not available for this checkpoint. NextStep-1.1’s documented strengths are its open-weight availability and distinctive autoregressive design; its practical suitability depends heavily on the hardware and pipeline surrounding it.

Availability and license

StepFun NextStep-1.1 is available as downloadable safetensors with custom inference code through Hugging Face and ModelScope. The official repository includes installation, model-download, inference, and training-related instructions. The Hugging Face model card lists the Apache 2.0 license.

Before commercial deployment, review the exact license notices for the checkpoint, repository, and dependencies, and test the model on the prompts, resolutions, and hardware relevant to your application. The model’s open-weight status provides deployment flexibility, but it does not guarantee a particular speed, image quality level, or supported configuration.