StepFun Releases Step 5 Preview Flagship Model, Open Weights on Oct 15
Key Highlights
Chinese AI lab StepFun released Step 5 Preview, a flagship base model built on a sparse multi-expert architecture with 600B total parameters and 27B active parameters. It supports a one-million-token context and accepts both text and vision input. On the Artificial Analysis Intelligence Index it scored 44, placing it among the top three openly usable models worldwide, with per-task cost claimed to be just one-eighth that of Claude Opus 5. The release is notable because it pairs a frontier-scale parameter budget with an explicit openness and cost message aimed directly at enterprises wary of closed APIs.
What It Does
Step 5 Preview is positioned as a general-purpose base model rather than a narrow vertical specialist. Its one-million-token context lets it ingest an entire book or a full codebase in a single session, while text-plus-vision input means it can describe images, read charts, and understand illustrated documents. For tasks that demand long-range reasoning and cross-document synthesis, this long-context, multimodal combination is simply more practical than short-context alternatives that must be fed information in small, lossy chunks. The vision capability also broadens the range of documents it can handle without external preprocessing.
Technical Details
The core idea of the sparse-expert design is large total, small active: the 600B total capacity handles knowledge breadth, while each inference call wakes only 27B active parameters, balancing capability against compute cost. The hard part of such architectures is routing — deciding which experts should be activated at the right moment. StepFun's choice to open the weights on October 15 effectively lays its training results open for community audit and secondary development, which is critical for building trust. Open weights also let researchers verify the claimed efficiency rather than taking the vendor's word for it.
Comparison With Peers
Among same-tier open or openly usable models, Step 5 Preview directly targets closed flagships like Claude Opus 5 with its 600B/27B MoE plus million-token context plus vision combination, using one-eighth the cost as its hook. Its rivals span both domestic open models such as DeepSeek and Qwen and the international open-source camp of Llama and Mistral. Whether it truly lands depends on post-release benchmarks and ecosystem growth, not the launch-day scores. The cost claim in particular will be tested once independent teams run their own workloads and publish real per-task spending.
Industry Implications
For domestic developers, another powerful open base model means a controllable, localizable alternative outside closed APIs, especially appealing to enterprises sensitive to data compliance and cost. Put simply, when top-tier capability starts arriving in an open plus low-cost form, the pricing logic of closed vendors comes under steady pressure. The October 15 weight release will be the key moment that tests whether this promise holds up, and it may accelerate adoption of open weights in production across regulated industries.
What to Watch
The signals to monitor are the actual open-weight quality after October 15, the maturity of the surrounding fine-tuning and inference tooling, and independent benchmark results on reasoning and vision tasks. If the community confirms the claimed cost-performance ratio, Step 5 Preview could become a serious default base for cost-sensitive production deployments in China, and a template for how open labs compete with closed flagships on economics rather than only on raw scores.
The Stakes
The bigger story is the economics of frontier AI. For two years the narrative has been that only a handful of well-funded labs could field flagship-grade models, and that openness meant accepting a quality gap. StepFun's explicit one-eighth-cost claim, if independently verified, challenges that assumption directly and gives cost-sensitive enterprises a credible alternative to closed APIs. The open-weight date is the moment of truth: until then the numbers are a promise. But the direction is clear — openness is moving up the quality curve while closing the price gap, and that pressure compresses the margin closed vendors can charge for comparable capability. For the global market, a strong open model from China also means the default assumption that the best weights come from the United States is no longer safe.
Bottom Line
Step 5 Preview is less a single product than a signal that the frontier is widening. Whether it ultimately beats closed flagships on every metric is almost secondary; what matters is that a credible, low-cost, open alternative now exists at this scale. The weeks after October 15 will determine if the community's independent tests match the launch claims. If they do, the practical default for many production deployments may shift from renting closed intelligence to hosting open intelligence — a change with consequences far beyond one model release, and one that rewards labs willing to let the community verify their math.
Reader Takeaway
For practitioners, the practical move is to wait for October 15 and then run their own workloads against the open weights before committing budget. Claims of cost advantage are only useful once measured on real tasks, and the community's independent numbers will matter more than the launch announcement.