StepFun Releases Step 5 Preview

StepFun (阶跃星辰) has released its agent-specialized flagship model, Step 5 Preview. The model adopts a sparse MoE architecture where only 27B parameters are activated per token out of a total of 600B, supporting a 1M token context and image/video input. It is currently available only via API and StepFun's own products, with an open-weight release scheduled for October 15.

Cost Efficiency and Intelligence

On the Artificial Analysis Intelligence Index, it scored 44 points, tying with Kimi K3 (Max) and trailing GLM-5.3 (Max) by 1 point. The core strategy is not competing for top performance, but reducing costs within a specific intelligence tier. The cost per task is approximately 1/3 ($0.72) of Kimi K3 ($2.00) and GLM-5.3 ($2.01), which StepFun labels as '-65% Cost'. API pricing is $1.00 for input and $2.70 for output (per 1M tokens).

Benchmark Performance

On its proprietary benchmark, StepCodeBench, it achieved an average of 49.0%, surpassing Kimi K3 (43.9%) and GLM-5.3 (40.2%). It showed particular strengths in bug fixing, feature implementation, and refactoring, but underperformed in performance tuning and documentation writing. In H100-based MLA kernel optimization experiments, it achieved 508 TFLOPS, outperforming Claude Opus 5 (493 TFLOPS). However, it scored lower than GLM-5.3 on some general agent benchmarks like Terminal-Bench v4, and acknowledged a gap in coding and long-running tasks compared to frontier models such as GPT-6 Astra.

Licensing and Usage

The license has not yet been finalized, but given that previous versions Step 3.5/3.7 Flash were under the Apache License 2.0, a similar open license is expected. Since the full 600B parameters must be loaded into memory, hardware requirements for self-hosting are high, but inference costs are expected to be low due to the small number of active parameters.