News
ReleasesOVHcloud unveils Olala mini, its first in-house LLMChinaA 9B model beats a 1T model on FinQA, trained with Crusoe Managed RLReleasesThe internal metaphysics of an LLMReleasesMapping Language Model Internal Representations to Physical Laws and Scientific ReasoningResearchCan China keep rogue AI in check? Meet the start-up testing the top modelsFrontierMicrosoft aims to challenge OpenAI, Anthropic with cheaper, specialized AI modelsOpen modelsMicrosoft rolls out new first-party models and updates to agentic AI continuous improvementFrontierMicrosoft unveils AI models for coding, cybersecurity as it takes on Anthropic, OpenAIFrontierGoogle is making it easier to opt out of Gemini AI model training – here’s howFrontierOpenAI Unveils New Astra AI Model After Scrapping Prior ReleaseReleasesOVHcloud unveils Olala mini, its first in-house LLMChinaA 9B model beats a 1T model on FinQA, trained with Crusoe Managed RLReleasesThe internal metaphysics of an LLMReleasesMapping Language Model Internal Representations to Physical Laws and Scientific ReasoningResearchCan China keep rogue AI in check? Meet the start-up testing the top modelsFrontierMicrosoft aims to challenge OpenAI, Anthropic with cheaper, specialized AI modelsOpen modelsMicrosoft rolls out new first-party models and updates to agentic AI continuous improvementFrontierMicrosoft unveils AI models for coding, cybersecurity as it takes on Anthropic, OpenAIFrontierGoogle is making it easier to opt out of Gemini AI model training – here’s howFrontierOpenAI Unveils New Astra AI Model After Scrapping Prior Release
EveryLLM
imgvidfnStep-5-Preview

Step-5-Preview

Agentic model with 600B parameters and 1M token context

Step-5-Preview is a large language model built by StepFun for complex agentic tasks. Released in 2026, it uses a sparse mixture of experts architecture to handle software engineering and financial analysis. The model supports text, images, and video inputs with a context window of up to one million tokens.

Share on WhatsApp

  • 600 billion total parameters with 27 billion active
  • Supports context windows up to 1 million tokens
  • Designed for agentic software engineering and finance
  • Accepts text, image, and video inputs
  • Deployable via vLLM, SGLang, or Transformers

What it is

Step-5-Preview is the flagship artificial intelligence model from StepFun, designed specifically for agentic workflows. It features a total parameter count of 600 billion, with 27 billion active during any given operation. This sparse design allows it to process up to one million tokens of context while maintaining efficiency. The system accepts text, images, and video as input. It is built to support sustained reasoning and tool use across complex professional tasks, particularly in software engineering and finance.

How it came about

The model was introduced by StepFun as their leading solution for professional knowledge work. Documentation from September 2026 highlights its performance on benchmarks such as GDPval-AA v2.1 and DeepSWE v1.1. The release includes specific deployment guides for frameworks like vLLM and SGLang, indicating a focus on practical integration for developers. The company positions it as a tool for high-level reasoning and iterative code verification.

Inside the model

The architecture relies on a sparse Mixture-of-Experts approach. This means only a subset of the neural network activates for each token, reducing computational load despite the large total size. It supports a context window and maximum output of up to one million tokens. The model integrates reasoning parsers and tool-call parsers to facilitate agent-like behaviour. It can be deployed using FP8 or BF16 precision on hardware such as Nvidia B300 or H200 GPUs, utilizing tensor, data, or expert parallelism.

What it can do

Step-5-Preview excels in repository-level software engineering, allowing it to navigate large codebases and coordinate file changes. It performs long-horizon agentic execution, which involves repeated tool calls, verification, and revision. In professional domains, it analyses long documents, charts, and multimedia to produce structured reports. The company claims particular strength in financial tasks that require combining quantitative reasoning with document analysis. Evaluation results are reported using high reasoning settings.

How to use it

The model weights are available for offline generation via the standalone Transformers implementation. Developers can also serve it through an OpenAI-compatible API using vLLM or SGLang frameworks. Deployment configurations are provided for various GPU setups, including single-node and multi-node clusters. The system supports both FP8 and BF16 checkpoints, with specific instructions for managing memory and parallelism strategies like MegaMoE on Blackwell architecture.

Questions people ask about Step-5-Preview

What is the parameter size of Step-5-Preview?

The model has 600 billion total parameters. However, it uses a sparse Mixture-of-Experts architecture, meaning only 27 billion parameters are active for each token processed. This design balances high capacity with computational efficiency.

What types of input does the model accept?

Step-5-Preview is multimodal. It accepts text, images, and video as input. This allows it to analyse charts, screenshots, and long documents alongside textual data for comprehensive professional knowledge work.

How long can the context window be?

The model supports a context window of up to one million tokens. This limit applies to the combined total of input and output tokens, enabling it to handle very large codebases or extensive document sets.

What hardware is recommended for deployment?

Deployment guides suggest using Nvidia B300 or H200 GPUs. Configurations vary from single nodes with four or eight GPUs to multi-node setups with sixteen GPUs, supporting FP8 and BF16 precision levels.

Good reading and sources

    Where this page comes from

    There is no English Wikipedia article on Step-5-Preview yet, so the facts on this page come from the sources listed above: huggingface.co/stepfun-ai/Step-5-Preview. Some are the subject's own pages, so its claims about itself are reported as its claims. EveryLLM wrote the text in its own words, with AI assistance, and compares it with the sources to make sure no passage is copied.

    It may contain mistakes, and figures change over time. Nothing here is investment or buying advice. If something is wrong, tell us. Last refreshed 2026-11-14.

    The weekly digest

    Every Monday: the model of the day, new pages on EveryLLM, and the conferences and anniversaries of the week ahead.