NextStep-1.1
NextStep-1.1 is a vision large model (VLM) developed by StepFunAI. It is a foundation model with a parameter scale of approximately 150.0B and is released under the Apache 2.0 license.
Parameters
150.0B
Context Window
License
Apache 2.0
Release Date
2025-12-24
API Pricing
API pricing for this model is not yet available
Strengths
- ・Large 150B parameter scale
- ・Advanced visual understanding capability
- ・Open Apache 2.0 license
Weaknesses
- ・High computational cost due to scale
- ・Resource consumption during inference
- ・Hardware requirements for operation
Use Cases
- ・Advanced image analysis
- ・Understanding and processing visual information
- ・Open-source visual AI development
Deep Analysis
Architecture
Autoregressive Image Generation
Continuous token generation
Training Steps
500K total
200K→500K from NextStep-1
Resolutions
256px + 512px
Multi-resolution training
Release Date
2025
ICLR 2026 Oral
Post-Training
NextStep-GRPO
Stabilized reinforcement learning
License
Open-source
Strengths
- ・ICLR 2026 Oral presentation
- ・SOTA autoregressive image generation with continuous tokens
- ・Addresses instability issues of original NextStep-1
- ・Significant improvement in image quality and text rendering
- ・Open-source research project
Weaknesses
- ・Research model, not production-ready
- ・Limited to image generation
- ・Requires significant compute for training
Competitor Comparison
| Model | Arena | SWE | GPQA | Price |
|---|---|---|---|---|
| NextStep-1 | - | - | - | Free (open) |
| DALL-E 3 | - | - | - | Paid |
| Stable Diffusion 3 | - | - | - | Free (open) |
NextStep-1.1 is StepFun's improved autoregressive image generation model, presented as an ICLR 2026 Oral paper. It addresses visualization failures from NextStep-1 through extended training (500K steps) and a stabilized post-training strategy (NextStep-GRPO).
Analysis generated: 2026-05-24