StepFun has introduced the Step 5 Preview model, a sparse Mixture-of-Experts architecture featuring 600 billion total parameters and 27 billion active parameters per token. The model utilizes a narrow-deep design with 92 Transformer layers to improve multi-hop reasoning capabilities. It supports a 1 million token context window, specifically designed for long-horizon agentic tasks.
The model is currently accessible via a hosted API and the StepFun platform, with open weights scheduled for release on October 15, 2026. Technical optimizations include MTP-3 speculative decoding, FP8 MoE, and KV-cache offload, which the company claims result in a 3x speedup for long-horizon reinforcement learning. Deployment requires approximately 1.2 TB of VRAM in BF16 format.
In performance evaluations, Step 5 Preview achieved scores of 67.7 on DeepSWE v1.1, 49.0 on StepCodeBench, and 80.5 on ProgramBench. Artificial Analysis assigned the model a score of 44 on its Intelligence Index, noting an output speed of 99.8 tokens per second. The model demonstrated capabilities in agentic experiments, including H100 kernel tuning and automated post-training tasks.
The research team documented the model performing 950 web fetches in a single agent action. Additionally, StepFun integrated Claude Code through its Step Plan feature. While the model shows competitive performance metrics, it currently trails GPT-6 Astra and Claude Opus 5 in standard coding benchmarks.
