Step 5 Preview Joins Tencent WorkBuddy — Free for Two Weeks

Tencent's all-scenario AI office platform WorkBuddy has announced the integration of StepFun's latest flagship model, Step 5 Preview, offering users exclusive free access for two weeks. This partnership embeds a 600-billion-parameter flagship foundation model directly into real workplace scenarios — spanning deep research, information retrieval, data analysis, and document and presentation generation. It reflects a broader shift in how large model capabilities are being deployed: moving beyond simple Q&A toward the continuous execution of complex, multi-step tasks.

Step 5 Preview: Balancing Performance and Cost with Sparse MoE

Step 5 Preview is StepFun's next-generation flagship foundation model. It uses a sparse MoE (Mixture of Experts) architecture with 600 billion total parameters, but activates only around 27 billion parameters per token. This design has become mainstream in the large model space — sparse activation lets a model retain the broad intelligence that comes with massive parameter counts while keeping inference cost and latency within acceptable bounds.

StepFun positions Step 5 Preview as optimized for AI coding, software engineering, professional knowledge work, and finance, emphasizing the best balance between flagship-level general intelligence, efficiency, and cost. In other words, the target isn't topping leaderboards in a lab — it's the productivity use cases that enterprises and developers are willing to pay for consistently.

Third-Party Benchmarks and Platform Data

Publicly available data suggests Step 5 Preview's performance holds up under scrutiny. On Artificial Analysis, a leading global LLM evaluation platform, it scores 44 on the Intelligence Index, placing it in the top three among open models worldwide. Even more telling are the real-world usage numbers: last week, Step 5 Preview topped the OpenRouter Trending chart and ranked second globally on the daily leaderboard by token consumption.

The OpenRouter usage ranking is particularly noteworthy — it's developers voting with their wallets. Reaching second place globally in token consumption means a large volume of real applications are actively calling this model, not just citing its benchmark scores. This combination of leaderboard performance and real-world usage data carries more weight than benchmark results alone.

---

What is sparse MoE? The core idea behind sparse MoE (Mixture of Experts) architecture is to split a model into multiple "expert" sub-networks. A gating mechanism (Router) dynamically selects a small number of experts to handle each input, rather than activating all parameters at once. In Step 5 Preview's case, while the total parameter count reaches 600 billion, each token only invokes around 27 billion parameters — making the actual compute comparable to a 27-billion-parameter dense model, while retaining knowledge capacity closer to a much larger model. Prominent examples of this architecture include Google's Switch Transformer and the Mixtral series. The main challenges with MoE are load balancing during training (preventing over-reliance on certain experts) and VRAM requirements at inference — even though compute drops, all expert weights still need to be loaded into memory, raising the hardware bar for deployment.

What is OpenRouter? OpenRouter is an API aggregation platform that gives developers access to hundreds of model endpoints — including GPT, Claude, Gemini, and various open-source models — through a single unified interface, billed by actual token usage. Its Trending chart and daily token consumption leaderboard reflect real developer choices in production projects, making them a strong market signal. Unlike benchmark rankings, which can be gamed through targeted optimization, OpenRouter usage data is much harder to manipulate. A model cracking the global top two in token consumption typically means it has earned developer trust through a strong combination of cost, performance, and reliability — and has been integrated into production-scale applications at meaningful volume.

---

How Model Capabilities Land in WorkBuddy's Agent Scenarios

Tencent WorkBuddy^Ad is positioned as an all-scenario AI office platform, with Agent (intelligent agent) capability as its core value proposition. With Step 5 Preview now integrated, users can leverage its reasoning and Agent capabilities to handle complex office tasks — deep research, information retrieval, data analysis, document generation, and slide creation — with support for sustained, multi-step task execution.

The key phrases here are "multi-step" and "sustained execution." Real office tasks rarely resolve in a single exchange. Producing an industry analysis report, for example, requires the model to first retrieve information, then analyze and organize it, and finally generate a formatted document or presentation. This demands end-to-end continuity across retrieval, analysis, and delivery — not just strong performance at any single point.

What This Partnership Means for Both Sides

This collaboration is mutually beneficial. For WorkBuddy, integrating a flagship model with standout reasoning and Agent capabilities helps raise the quality and efficiency of task execution. For StepFun, it opens an important gateway into real workplace scenarios.

A common challenge for model providers is moving beyond API calls toward genuine use-case depth. Offering a model endpoint alone makes it easy to get trapped in price wars and commoditization. By partnering with a platform like Tencent WorkBuddy — which has a large base of enterprise office users — StepFun can accumulate real usage data and feedback from high-frequency, high-value workplace contexts. That directly supports future model iteration and commercialization.

---

AI Agent vs. standard LLM chat — what's the difference? The fundamental distinction is autonomous planning and tool use. In a traditional chat interaction, the model passively responds to a single prompt. In Agent mode, the model breaks a complex goal into sub-tasks, sequentially calls external tools — search, code execution, file read/write, API requests — and dynamically adjusts subsequent steps based on intermediate results, until a final deliverable is produced. This depends on the model's ability to reason across long task chains and to precisely follow tool-calling formats (such as Function Calling). Common Agent frameworks include ReAct (Reasoning + Acting loop) and Plan-and-Execute paradigms. In office settings, Agent capability directly determines whether a model can reliably complete end-to-end tasks like "retrieve information and produce a full report" — not just shine at single-step text generation.

---

The Logic Behind Two Weeks of Free Access

Offering exclusive free access for two weeks is a classic user acquisition and scenario validation play. For users, it's a zero-cost window to experience flagship model Agent capabilities firsthand. For the platform and model provider, it's an opportunity to gather substantial real-world usage data and test actual model performance on office tasks in a compressed timeframe.

Zooming out, this partnership reflects a new phase in the competition among Chinese large model providers. As parameter counts and benchmark scores converge, the battleground is shifting toward who can most quickly translate model capability into usable productivity tools — and who can secure more real-world deployment scenarios. StepFun's decision to partner with Tencent WorkBuddy and go deep on office workflows is a concrete move in exactly that direction.

For users following AI office tools, these two weeks of free access are worth treating as a hands-on evaluation opportunity — a chance to put Step 5 Preview through its paces on complex document processing, data analysis, and similar real tasks.