JieYue Step 5 is officially released, and will be open-sourced on October 15.

Beating AI News: StepStar officially launches its flagship model Step 5 Preview. Built on a sparse Mixture-of-Experts (MoE) architecture, the model has a total of 600 billion parameters, with only 27 billion activated per inference run. It supports a 1 million-token context window and text-image input. The model’s API and Studio are now fully open, with full model weights scheduled for release on October 15. Artificial Analysis (AA) completed pre-evaluations earlier. Step 5 Preview scored 44 points on the Intelligence Index, matching Kimi K3 Max. AA currently measures its output speed at ~100 tokens per second, with a task cost of $0.71, compared to Kimi K3 Max’s $2 per task. The developer highlighted the model’s long-task capability: StepStar enabled Step 5 Preview to autonomously optimize a set of H100 GPU cores for up to 24 hours straight. The model independently modifies code, runs tests, compares results, and iterates. After roughly 22 hours, it achieved 508 TFLOPS, while Claude Opus 5 reached 493 TFLOPS in the same experiment. In another 24-hour trial, Step 5 Preview designed its own training data, lifting Qwen3-30B-A3B’s accuracy on AIME24 from 53.3% to 60% — matching Claude Opus 5, while consuming fewer labeled tokens.