StepFun's Step 5 Preview arrives as a 600B-parameter sparse mixture-of-experts activating roughly 27B per token, paired with a 1M-token window and a narrow-deep 92-layer stack, with weights slated for open release. The headline number matters less than the design intent: this is a model built for long-horizon agents, not chat.
Globally, the move sharpens a trend that closed-lab pricing has tried to obscure. Sparse MoE lets a lab claim frontier-scale capacity while keeping inference cost closer to a mid-sized dense model, and a deep-narrow architecture with a million-token context is a direct bet that the next value pool is sustained, multi-step agentic work rather than single-turn responses. When those weights are open, the competitive pressure lands on two fronts. First, it erodes the moat around long-context reasoning, which OpenAI, Anthropic, and Google have monetized at premium API tiers. Second, it hands regulated and cost-sensitive enterprises a self-hostable path to capabilities they previously had to rent. The risk to watch is practical: 1M-context throughput is expensive to serve, and open weights don't erase the memory and interconnect bill. The real differentiator becomes who can run this economically, not who can download it.
For Japan, the strategic implication is unusually direct. Data-residency expectations in finance, government, and healthcare have made pure API dependence on US frontier models a persistent governance headache. An openly licensed long-context MoE gives Japanese enterprises and their SIer partners a credible on-premise or sovereign-cloud option, though the license terms and any provenance concerns tied to a China-origin model will need careful legal review before regulated sectors touch it.
For SIers such as NTT Data, Fujitsu, and NRI, this reframes the offering from 'model access' to 'model operations.' The margin sits in fine-tuning, retrieval integration, and building the serving stack that makes a 600B MoE affordable on Japanese infrastructure. That plays to an integration strength while exposing a real gap in GPU capacity and MLOps depth. For RPA-heavy operations, a genuine long-horizon agent threatens the brittle, rule-based automation that still dominates Japanese back offices. The teams that pair these open weights with disciplined evaluation and human oversight will convert the disruption into a services opportunity rather than a displacement.
