StepFun has introduced Step 5 Preview, its latest foundational AI model specifically engineered for sophisticated agentic tasks across software engineering, professional knowledge management, and financial analysis. A primary focus of this release is delivering advanced intelligence at a significantly reduced operational expense, positioning Step 5 Preview as a cost-efficient alternative in the competitive AI landscape.
Deployment and Future Plans
Step 5 Preview is currently accessible via a hosted API and within the StepFun platform. Open weights for the model are slated for release on October 15, 2026. Considering its substantial 600 billion parameters, future self-hosting will necessitate robust multi-GPU server infrastructure, requiring significant memory.
Technical Overview
Step 5 Preview features a sparse Mixture-of-Experts (MoE) architecture with approximately 600 billion total parameters, yet only about 27 billion are actively engaged per token (roughly 4.5% of its overall weight). Key specifications from official documentation include a 1 million-token context window, multi-modal input (text, image, video) with text output, selectable reasoning intensity, and support for streaming, tool calling, and JSON modes. StepFun reports the model coordinated 950 web fetches in a single agent action during research tasks.
Architectural Design and Training Innovations
The model employs a "narrow and deep" architecture, stacking 92 Transformer layers. This design prioritizes extended pathways for implicit multi-hop reasoning, deemed crucial for long prefill operations common in agentic workflows. Training for Step 5 Preview utilizes on-policy, long-horizon reinforcement learning. Advanced techniques cited include bit-wise train and inference alignment across MoE routing, MTP-3 speculative decoding, FP8 MoE implementation, and KV-cache offloading. These innovations reportedly contribute to a threefold or greater end-to-end speed enhancement for long-horizon reinforcement learning.
Performance Benchmarks
StepFun has provided internal performance data, with Step 5 Preview operating at "High" effort. While benchmark scores were reported for areas like finance and coding, competitor models often maintained a lead. However, Step 5 Preview demonstrated strong performance in two 24-hour agent experiments: one where it optimized an H100 kernel, surpassing a rival's result, and another where it boosted a large language model's performance via automated post-training.
Artificial Analysis independently scored Step 5 Preview at 44 on its Intelligence Index, significantly above the median for comparable reasoning models. Its API output rate was measured at 99.8 tokens per second.
Pricing Structure
StepFun's API list prices per 1 million tokens are $1.00 for input (cache miss), $0.05 for input (cache hit), and $2.70 for output. While Artificial Analysis reports median prices for similar models at $1.88 input and $10.00 output, Step 5 Preview's verbosity is a consideration. It generated a significantly higher volume of output tokens compared to the median, potentially impacting overall cost efficiency despite lower per-token rates.
Key Features Summary
- StepFun’s Step 5 Preview is a 600B-total, 27B-active Mixture-of-Experts model.
- It offers a 1 million-token context window with multi-modal input.
- API pricing stands at $1.00 per 1 million input tokens and $2.70 per 1 million output tokens.
- Artificial Analysis scored the model 44 on its Intelligence Index.
- Open weights are scheduled for October 15, 2026.
This article is a rewritten summary based on publicly available reporting. For the original story, visit the source.
Source: MarkTechPost
