Urgent.News

What's breaking now, across hundreds of outlets.

AI

Step 5 Preview, a 1M-context MoE from StepFun, shows up on OpenRouter

Step 5 Preview, a massive 1 million-context model developed by StepFun, has emerged on OpenRouter. This model, designed for agentic tasks, is based on a sparse Mixture-of-Experts architecture with 27B active experts out of a total of 600 billion parameters. Step 5 Preview excels in software engineering and professional knowledge work, particularly in the finance domain.

It is tailored for long-term tasks involving large codebases and documents, utilizing tools and refining results across several steps. In the event of an error from an upstream provider, the system can gracefully recover by switching to another functional provider, assuming the request's filters permit such a maneuver. Detailed uptime data for each provider is available programmatically through the Endpoints API. Further insights into load balancing and customization options are available on StepFun's platform.

Meanwhile, StepFun has also unveiled Step 3.7 Flash, a cutting-edge high-efficiency multimodal Mixture-of-Experts model. This model combines a 196 billion-parameter language backbone with a vision encoder capable of native image and video comprehension. Roughly 11 billion parameters of the model are activated per token. Step 3.7 Flash supports a 256K context window and offers selectable reasoning levels (high, medium, low), enabling users to balance speed, cost, and depth of reasoning according to their needs.

The model is particularly suited for coding, agentic workflows, structured outputs, and long-context productivity tasks.

In contrast, Step 3.5 Flash represents StepFun's most capable open-source foundation model. This model is also based on a sparse Mixture of Experts architecture and selectively activates only 11 billion parameters of its 196 billion-parameter total per token. Despite its efficiency, Step 3.5 Flash is a powerful reasoning model, demonstrating exceptional speed even when handling long contexts.

Rewritten by urgent.news; may contain errors — check the original.

Source: openrouter.ai

Read the original at openrouter.ai →

Report a problem with this story

More in AI

Google unveils Gemini AI agent to handle workplace tasks

The Gemini agent can write and run code, integrate into apps like Gmail, and help workers knock out everyday workplace tasks, Google said.

  • Google introduces Gemini AI agent to streamline workplace tasks
  • Gemini can write code, generate images, and interact with Google and third-party apps
  • Workers can delegate tasks to Gemini, which builds memory and skills over time

More from Thursday 8 October →