StepFun drops Step 5 Preview, a 600B MoE model built for agents

Now that the era of raw benchmark bragging has cooled down, builders care about a much simpler question: can your model actually survive a multi-step agent workflow without bankrupting you?

This week, Chinese AI lab StepFun stepped into that arena with the launch of Step 5 Preview, its latest flagship multimodal model tailored specifically for agentic work. It is already accessible through the StepFun Open Platform.

600B under the hood, 27B on the meter

Step 5 Preview leans heavily into a smart Mixture-of-Experts (MoE) setup.

The entire network contains a massive 600 billion parameters, but during inference, only 27 billion parameters are active for any given token.

That dynamic routing delivers the depth of an frontier model while drastically cutting down latency and computational overhead.

It also pairs that architecture with a 1 million token context window. That means you can dump entire codebases or lengthy technical dossiers straight into prompt memory without losing track of details.

Running neck-and-neck with the best, at one-third the price

StepFun clearly tuned this model for complex knowledge and coding tasks.

On the DeepSWE coding benchmark and various knowledge-work evaluations, Step 5 performs on par with or ahead of top open-weight competitors like GLM 5.3 and Kimi K3.

According to rankings from Artificial Analysis, Step 5 sits just a single point behind GLM 5.3.

The real punchline is the price tag: Step 5 is nearly three times cheaper to run than GLM 5.3, and significantly cheaper than the dominant closed-source APIs.

Because it handles vision natively, it fits right into modern vibe coding setups. For instance, developers can feed it an image and prompt it to generate and render a 3D asset directly. It also handles end-to-end financial data synthesis with minimal hand-holding.

StepFun offers an OpenAI-compatible API and a browser-based Studio playground right now, with a formal public release scheduled for October 15th.

Why it matters

For builders creating autonomous agents, this release solves an acute pain point. Multi-turn reasoning loops burn through tokens at terrifying speeds, and having an frontier-tier coding model that cuts API bills by nearly two-thirds makes agent projects far more economically viable.

For the broader AI ecosystem, it proves that the frontier race is no longer just about gross parameter counts. Pushing 600B capability into a 27B active footprint while driving costs to rock-bottom levels sets a punishing standard for competing labs.

In our view, StepFun played this smart by prioritizing practical developer economics over marketing hype. That said, it is still labeled as a Preview. Whether its agentic reliability holds up when faced with messy real-world tool execution will be the true test once October 15th rolls around.