Step 5 officially released, open-sourced on October 15

Beating AI News Flash: StepFun officially releases its flagship model Step 5 Preview. It adopts a sparse MoE architecture with 600B total parameters, activating only 27B per pass, and supports a 1 million Token context and image-text input. The API and Studio are now fully open, and the complete model weights will be released on October 15.

Artificial Analysis had already completed its evaluation in advance. Step 5 Preview scored 44 on the Intelligence Index, on par with Kimi K3 Max. AA currently measures its output at approximately 100 Token/s, with a cost of $0.71 per task; Kimi K3 Max is $2.

The official highlight is its ability to run long tasks. StepFun had Step 5 Preview autonomously optimize a set of H100 GPU kernels for up to 24 consecutive hours. The model would modify code, run tests, compare results, and continue iterating on its own. After about 22 hours, it reached 508 TFLOPS, compared with 493 TFLOPS for Claude Opus 5 in the same experiment.

In another 24-hour experiment, Step 5 Preview designed its own post-training data, raising Qwen3-30B-A3B's accuracy on AIME24 from 53.3% to 60%, on par with Claude Opus 5, while consuming fewer annotation Tokens.