NextStep 1 Large
NextStep 1 pairs a 14B autoregressive model with a 157M flow-matching head. It learns discrete text tokens and continuous image tokens under a next-token-prediction objective.
Understand what the model predicts
The release uses continuous image representations alongside discrete language tokens. The inference example loads a custom pipeline and a VAE rather than a text-only chat model.StepFun NextStep-1-Large model card
Treat decoding as part of the model
Practical guidance. Changing the image decoder or sampling settings can change the output even when the language backbone stays fixed. Preserve the complete pipeline when reproducing a result or evaluating a fine-tune.
Specifications & access
| Detail | Value and conditions |
|---|---|
| Access | Downloadable |
| Model identifiers | NextStep-1-Large |
| License | Apache 2.0.StepFun NextStep-1-Large model card |
| Architecture | Autoregressive backbone with a smaller flow-matching output head.StepFun NextStep-1-Large model card |
Related models
- NextStep 1 Large Edit
StepFun · NextStep
NextStep 1 Large Edit adapts the NextStep image pipeline to an input image and an editing instruction.
Launch date not confirmedDownloadable - NextStep 1 Large Pretrain
StepFun · NextStep
An image-generation pretraining checkpoint combining a 14B autoregressive backbone with a 157M flow-matching head.
Launch date not confirmedResearch releaseDownloadable - NextStep 1.1
StepFun · NextStep
An autoregressive image-generation model trained with flow-based reinforcement learning.
Launch date not confirmedDownloadable