StepFun releases Step 5 Preview for long-horizon agent work
StepFun's new sparse MoE model holds about 600B total parameters, activates roughly 27B per token, and is set to get open weights on October 15, 2026.
StepFun has released Step 5 Preview, its new flagship model aimed at agentic workloads: software engineering, professional knowledge work and finance. The company's central claim is cost, saying the model reaches comparable intelligence at a substantially lower cost per task.
Step 5 Preview is a sparse Mixture-of-Experts model with roughly 600B total parameters and about 27B activated per token, close to 4.5% of the weights. The context window is 1M tokens. Instead of widening the network, StepFun went narrow and deep, stacking 92 Transformer layers according to Pandaily. The research team argues deeper stacks give longer paths for implicit multi-hop reasoning, which matters during long prefill while an agent searches, runs code and reads tool returns.
Training and measurements
Training relies on on-policy, long-horizon reinforcement learning, with bit-wise train and inference alignment across MoE routing. Listed techniques include MTP-3 speculative decoding, FP8 MoE and KV-cache offload, and StepFun reports more than 3x end-to-end speedup for long-horizon RL. On coding it reports 67.7 on DeepSWE v1.1, 49.0 on its own StepCodeBench, and 80.5 on ProgramBench. By StepFun's own account, GPT-6 Astra and Claude Opus 5 stay ahead on all three.
Two 24-hour agent experiments are also described. In one, the model tuned an H100 kernel to 508 TFLOPS against 493 for Claude Opus 5. In the other, automated post-training lifted Qwen3-30B-A3B on AIME24 from 53.3% to 60%. StepFun says the model coordinated 950 web fetches in a single agent action.
Artificial Analysis scores Step 5 Preview at 44 on its Intelligence Index, where the median for reasoning models in a similar price tier is 24, and measured output at 99.8 tokens per second on StepFun's API. One caveat: the index run produced 160M output tokens against a 92M median, so verbose reasoning eats part of the per-token savings.
The model is available through a hosted API and the StepFun platform, with open weights announced for October 15, 2026. Details are in the MarkTechPost report.
