Breaking

StepFun Launches Step 5 Preview, a 600B MoE Model for Agentic Work

October 1, 2026 at 15:46 EDT

Foundation Models
AI Agents
Software Dev & Coding

StepFun has introduced Step 5 Preview, a new flagship reasoning model aimed at agentic coding, professional knowledge work and finance, with 600 billion total parameters, 27 billion active parameters per token and a one-million-token context window. The launch puts another Chinese AI lab into the cost-performance race around large sparse models, though the strongest claims still need to be separated from what is independently measured.

Sep. 20, 2026 · StepFun

Step 5 Preview enters the agent race

A 600B-parameter sparse MoE model activates 27B parameters per token, targets long-horizon agentic work, and lists API pricing at $1 input and $2.70 output per million tokens.

600B

total parameters

27B

active per token

context window

$2.70

output per 1M tokens

Sparse compute is the cost argument

600B

total capacity

27B

active compute

The model uses about 4.5% of its parameter pool per token, shown here as one active block against twenty scaled capacity blocks.

Why teams are interested

Long-context coding agents, financial research and document-heavy workflows can test a 1M-token model at relatively low list prices.

What remains unsettled

The release is still a preview, several benchmark claims are vendor-reported, and high verbosity can raise real output cost.

The next checkpoint is the weight release

API

Available now as a hosted model.

Weights

Scheduled for October 15, 2026.

Proof points

Independent scoring is strong, but full reproducibility is still pending.

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.