Step 5 Preview gets one-week free trial on Nous Portal
Nous Research launches a 7-day free trial for Step 5 Preview. The MoE model features 600B total parameters, active 27B, and a 1M context window.
Developed by Chinese startup StepFun, the model combines 600 billion total parameters with 27 billion active per token and a 1-million-token context window.
Nous Research launched a one-week free trial of Step 5 Preview on its Nous Portal on Wednesday (October 9, 2026). Developed by Beijing-based startup StepFun, the system operates on a sparse Mixture-of-Experts (MoE) architecture that totals 600 billion parameters while activating only 27 billion for any given token.
The trial allows developers and researchers to evaluate the model’s capabilities without inference costs directly through the Nous Research interface. The seven-day promotional window serves as a hands-on preview ahead of the open weights release scheduled for mid-October.
Inside the Step 5 Preview architecture
Traditional dense models activate every parameter across the entire network for every single calculation. The MoE architecture takes a different path: think of a specialized workplace staffed by dozens of domain experts, where only two or three individuals are assigned to each specific query. In Step 5 Preview, the 600 billion parameters are distributed across 92 layers, but only roughly 4.5% of them participate in generating each token.
This selective activation drastically reduces hardware overhead and accelerates latency without sacrificing overall reasoning breadth. The system also supports native multimodal inputs, accepting text, images, and video. Its context window handles up to 1 million tokens, alongside a maximum continuous output of 64,000 tokens. In practice, that capacity allows developers to pass entire code repositories, technical manuals, or full video recordings in a single call without manual chunking.
Benchmark correction on the Hermes Index
Shortly after opening public access, Nous Research revised its internal benchmark figures on social media. In an initial post on X, the team reported a score of 33.89 for Step 5 Preview on the Hermes Index, which would have tied the model with OpenAI’s GPT-6 Luna. Minutes later, the lab published an official correction, adjusting the verified score down to 32.83.
Even with the downward revision, the final result keeps StepFun’s model competitive at the higher end of the leaderboard. It ranks directly behind GPT-6 Luna (33.89) and GLM 5.3 Flash (34.95). The central draw for independent developers is computational efficiency: reaching that benchmark tier while firing only 27 billion active parameters makes the model considerably less expensive to serve than dense rivals in the same class.
Commercial API and open-weight plans
Founded in Beijing by former researchers from Microsoft Research Asia and SenseTime, StepFun has offered commercial API access since September 20. Pricing sits at $1.00 per million input tokens and $2.70 per million output tokens. The company has also set up temporary free promotional access through coding assistant tools such as Cline and Kilo Code.
The wider milestone arrives on October 15, 2026, when StepFun plans to publish the model’s open weights. What remains unannounced are the exact licensing terms: the startup has not yet clarified whether the weights will carry a permissive license or enforce commercial caps based on monthly active users, a distinction that frequently determines enterprise adoption for Chinese open-weight models.