StepFun launches Step 5 Preview with 600B parameters, 1M context, and open weights coming October 15
StepFun has launched Step 5 Preview, a new flagship foundation model designed for long-running agentic work, software engineering, professional knowledge tasks and financial analysis.
The model uses a sparse Mixture-of-Experts architecture with 600 billion total parameters and 27 billion parameters activated per token, giving it a much smaller active computational footprint than its headline parameter count suggests.
Step 5 Preview supports a 1 million-token context window, native text and image input, extended reasoning and tool-based workflows.
API access is available from September 20, while StepFun says the model’s open weights will be released on October 15, 2026.
The distinction is important for developers evaluating the launch today: Step 5 Preview can already be called through StepFun’s infrastructure, but it is not yet an open-weight model available for independent deployment.
Independent evaluation by Artificial Analysis currently assigns the model an Intelligence Index score of 44, while listing API pricing at $1.00 per million input tokens and $2.70 per million output tokens.
··········
STEP 5 PREVIEW USES 600B TOTAL PARAMETERS WHILE ACTIVATING ONLY 27B PER TOKEN.
The sparse MoE architecture is designed to separate total model capacity from the amount of computation required for each generated token.
Step 5 Preview contains 600 billion total parameters, but only 27 billion are active for each token.
Data Studios calculates an active-to-total ratio of: 27B ÷ 600B = 4.5%.
In practical terms, approximately 95.5% of the model’s parameters are not activated for an individual token, although different expert subsets can be selected as the model processes different inputs.
This is the basic economic logic behind sparse Mixture-of-Experts systems. A dense 600B model would potentially need to involve essentially the entire parameter set during each forward pass, while a sparse MoE model instead contains a much larger collection of specialized parameter groups and routes each token through only a limited subset of them.
The architecture can therefore increase total representational capacity without increasing inference compute proportionally to the headline parameter count.
........
Specification | Step 5 Preview |
Total parameters | 600B |
Active parameters per token | 27B |
Active share of total parameters | 4.5% |
Inactive share per token | 95.5% |
Architecture | Sparse Mixture-of-Experts |
Context window | 1M tokens |
Native input | Text + images |
Output | Text |
API availability | Available |
Open weights | Scheduled for October 15, 2026 |
........
The 27B activation figure does not mean Step 5 Preview behaves like a conventional dense 27B model. The larger 600B parameter pool gives the routing system access to many more specialized experts across different tokens and tasks.
It also does not make deployment equivalent to hosting a 27B model. The complete weight set still has to be stored and distributed across hardware, meaning memory requirements, interconnect bandwidth and routing efficiency remain important when StepFun releases the weights.
The architectural advantage is therefore primarily reduced active computation relative to total capacity, not elimination of the infrastructure required to hold a 600B model.
··········
THE 1M-TOKEN CONTEXT TARGETS LONG-HORIZON AGENTIC WORK RATHER THAN SINGLE-PROMPT CHAT.
StepFun is positioning the model around workloads that accumulate large amounts of state, documentation, code and tool output over extended execution periods.
A 1 million-token context window allows a single model interaction to contain substantially more information than conventional chatbot conversations.
For software-engineering agents, that context can include repositories, specifications, issue histories, execution logs, test output and previous attempts at solving a problem. For financial or professional research, it can contain large collections of filings, reports, source documents and intermediate analysis.
Long context alone does not guarantee successful long-running agent behavior. An agent must still identify relevant information inside the window, preserve the correct state across many reasoning steps, use tools reliably and recover from failed intermediate actions.
This is why StepFun is emphasizing sustained execution alongside context length. The model is intended not merely to accept a large prompt, but to remain useful while a workflow expands through repeated cycles of reasoning, tool execution, observation and correction.
A representative agentic sequence can be expressed as: objective → context inspection → plan → tool call → result → state update → next action → verification → final output.
As this loop continues, context can accumulate rapidly. A million-token window gives the system more room to preserve that history before information has to be removed, summarized or externalized into another memory system.
The trade-off is computational. Large context windows increase prefill workloads and can materially affect latency and cost when applications repeatedly send hundreds of thousands of tokens.
For production systems, developers therefore still need retrieval, caching and context-selection strategies rather than treating the full 1M window as something that should be filled on every request.
··········
API PRICING PUTS STEP 5 PREVIEW INTO A DIFFERENT COST CLASS FROM MANY FRONTIER FLAGSHIPS.
The model is being positioned around capability per unit of inference cost rather than parameter scale alone.
Artificial Analysis currently lists Step 5 Preview at $1.00 per million input tokens and $2.70 per million output tokens.
It also reports a 95% prompt-cache discount, which would reduce repeated cached input to approximately $0.05 per million tokens under the measured pricing structure.
This matters particularly for agents because long-running workflows frequently reuse large portions of stable context. A coding agent may repeatedly send repository instructions, system prompts and project documentation, while a financial-research agent may repeatedly reference the same collection of filings or reports.
Caching those repeated tokens can materially change the economics of high-context applications.
........
Cost dimension | Reported Step 5 Preview pricing |
Input | $1.00 / 1M tokens |
Output | $2.70 / 1M tokens |
Cached input with 95% discount | ~$0.05 / 1M tokens |
Artificial Analysis Intelligence Index | 44 |
Context | 1M tokens |
Active parameters | 27B per token |
........
A simple workload illustrates the pricing structure. Suppose an agent processes 500,000 new input tokens and produces 50,000 output tokens.
At the listed prices, the uncached token cost would be 0.5 × $1.00 = $0.50 input and 0.05 × $2.70 = $0.135 output, for a total of approximately $0.635, excluding any additional platform charges or differences in production billing.
If most of a long prompt is reused through caching, the effective cost can fall further.
StepFun is explicitly presenting this combination of intelligence and cost as an improvement in the model’s Pareto frontier: increasing capability without moving proportionally upward in inference expense. That remains partly a vendor positioning claim.
Real application economics depend on output length, reasoning behavior, cache reuse, latency, tool calls and how frequently an agent has to retry or correct unsuccessful actions.
··········
THE OCTOBER 15 WEIGHT RELEASE WILL DETERMINE HOW OPEN THE MODEL ACTUALLY BECOMES.
API availability and open-weight availability are separate milestones, and the second one cannot yet be evaluated completely.
Step 5 Preview is usable through API access now, but StepFun says the weights will arrive on October 15.
Until that release occurs, developers cannot fully evaluate important deployment questions such as memory requirements, quantization behavior, multi-node inference efficiency, fine-tuning practicality and the licensing conditions attached to the weights.
The exact license will be particularly important. “Open weights” can describe substantially different levels of practical freedom depending on whether commercial deployment, modification, redistribution or derivative models are permitted.
Hardware requirements will also be significant. The model activates only 27B parameters per token, but the complete 600B parameter pool still has to reside somewhere in the inference system.
At BF16 precision, 600B raw parameters would correspond to roughly 1.2 TB of weight data before additional runtime overhead, providing an order-of-magnitude indication of why distributed deployment will remain relevant even with sparse activation.
Quantization could reduce that memory footprint considerably, but actual efficiency will depend on StepFun’s released implementation and supported inference stack.
The most useful way to interpret Step 5 Preview today is therefore as an API-accessible 600B sparse frontier model with unusually low active parameter count, very long context and aggressive inference pricing.
The October 15 release is the next technical checkpoint. At that point, the model can be judged not only on hosted benchmarks and API economics, but also on whether its sparse architecture translates into practical self-hosting, customization and deployment advantages for developers who want direct control over the weights.
··········
FOLLOW US FOR MORE.
·····
DATA STUDIOS
·····
[datastudios.org]
