AI development company StepFun has launched Step 5 Preview, a sparse Mixture-of-Experts foundation model featuring 600 billion total parameters and activating 27 billion parameters per token. Capable of processing multimodal inputs—including text, images, and video—the model offers a 1 million-token context window designed to support long-horizon agentic operations. Commercial API access is currently live at a rate of $1.00 per million input tokens and $2.70 per million output tokens, as reported by MarkTechPost. Additionally, the company has announced plans to release open weights for the model on October 15, 2026.
Targeting Complex Long-Horizon Workflows
The technical configuration of Step 5 Preview reflects an industry trend toward building systems tailored for extended autonomous task execution. By using a sparse Mixture-of-Experts architecture, the model maintains a high total parameter count while keeping active compute per token relatively low. StepFun is directing this capability toward complex domains such as software engineering, financial research, and professional knowledge work. In these areas, large context windows are essential for evaluating full code repositories, lengthy regulatory filings, and complex historical datasets without relying heavily on aggressive context truncation.
Missing Benchmarks and Open Questions
Despite the released architectural specifications, several critical parameters remain unannounced. The initial disclosure did not specify standardized performance benchmarks, leaving its actual accuracy and coding efficacy relative to competing frontier models unknown. Furthermore, details regarding training dataset provenance, safety alignment procedures, and context retrieval precision across the full 1 million tokens were not provided. The announcement also did not clarify the specific licensing terms that will apply to the scheduled October 2026 open-weights release, nor whether hardware requirements for hosting the 600-billion-parameter model locally will be practical for non-enterprise users.
As developer adoption proceeds through the public API, real-world benchmarks will determine whether the model fulfills its operational goals for long-horizon work.
Lead image: Brett Sayles / Pexels
