Step 5 Preview Tops OpenRouter Trending: 600B-Parameter Sparse MoE, Official Open-Source Date Set for October 15
According to a Leiphone report dated October 10, 2026, StepFun's Step 5 Preview topped OpenRouter's Trending list and ranked second globally on the daily list, using a sparse MoE architecture with 600 billion total parameters and about 27 billion activated per token; the company said it would open-source the model on October 15, and its Artificial Analysis composite intelligence index is 44.
What changed
According to a Leiphone report dated October 10, 2026, StepFun's latest flagship base model, Step 5 Preview, reached the top of OpenRouter's Trending list and ranked second globally on the daily list by total tokens processed per day. The report said official information indicated the model would be formally open-sourced on October 15. The architecture details given in the report are: a sparse MoE design with 600 billion total parameters, about 27 billion parameters activated per token, and optimizations for AI coding, software engineering and professional knowledge work, and finance scenarios. The report also cited the Artificial Analysis leaderboard, saying its composite intelligence index is 44 and that it ranks in the global open-source top three. In addition, AI researcher and DAIR.AI co-founder Elvis Saravia, after trying the model, said its performance was impressive, that its capabilities could be compared with models such as GLM-5.3 and Kimi K3, that its pricing is competitive, that it sits in a cost-versus-capability trade-off range, and advised developers to connect it to their own commonly used coding agents. The above rankings, benchmark scores, capability comparisons and pricing assessments all come from the report and its citations; they are relayed information rather than results independently verified by this publication.
What it means for developers
For developers, this news needs to be split into two separate matters. The first is an API-side fact that is usable now: the model has already generated considerable real call volume on OpenRouter, showing that its access path on the aggregator platform works; teams working on coding agents, code completion, or financial text processing, which are high-token-consumption scenarios, can include it in the same OpenAI-compatible calling layer for A/B comparison, using the same prompt set to measure completion rate, retry rate, and end-to-end latency. The second is something that has not yet happened: the company said it would open-source the model on October 15, but the report only gives a date and does not say what the weight license will be, whether commercial terms will be attached, or whether inference code and quantized weights will be released at the same time. Open weights are not the same as an open-source license; whether self-deployment is possible, whether commercial use is allowed, and whether distillation is permitted all have to wait until the license text lands. Teams should not plan their architecture in advance on the assumption that it can be self-hosted. In addition, Elvis Saravia's assessment of a "cost versus capability trade-off" points to the model's price competitiveness, but specific API pricing changes; readers should rely on real-time pricing pages from vendors and platforms rather than citing figures from around the report date for long-term budgeting. The sparse MoE structure also means that self-hosting has significantly higher VRAM and parallelism thresholds than a small model with the same activated parameter count: 27 billion activated may sound acceptable, but 600 billion total parameters determine the weight magnitude and the communication overhead of expert routing, so local deployment is more likely to fall into multi-GPU or cluster-level solutions.
Integration and evaluation
Recommendation one: if the team's bottleneck is the unit cost of coding or software engineering tasks, OpenRouter can first be used as a quick trial channel. The precondition is that these tasks allow data to be forwarded through a third-party aggregator platform; if there are data cross-border or customer compliance constraints, this channel does not hold, and evaluation should wait until the official API or weights are released. Evaluation criteria can be set as: on the same evaluation set as the existing main model, the pass rate is not below the baseline, the average token consumption per task does not rise significantly, and p95 latency is within an acceptable range. Recommendation two: if the team truly cares about self-hosting, it should not make deployment decisions based on the statement that it will be open-sourced on October 15, but should break this step into preconditions to be verified: whether the weights can be downloaded, whether the license allows the target commercial scenarios, whether the vendor provides recommended inference frameworks and quantization schemes, and whether 600 billion total parameters can run at acceptable throughput on the team's own hardware. Only after the license text and weights both land, and a measurable throughput figure can be run on a single GPU or the existing cluster, should the team decide whether to switch production traffic over; the evaluation criteria are the amortized self-hosting cost per million tokens and online consistency, meaning the degree of output quality degradation after quantization. Recommendation three: treat the report's "composite intelligence index of 44, open-source top three" as a clue to be rechecked rather than a conclusion, and supplement testing with your own private set before adoption, especially codebase understanding under long context and structured inputs such as financial tables, because these scenarios are closer to the optimization directions described in the report and are also the most likely to expose shortcomings beyond official evaluations.
Limits and open questions
The report was published on October 10, 2026, earlier than the claimed October 15 open-source date, so whether the open-sourcing is delivered on schedule, under what license it is released, and whether the weights are complete cannot be confirmed from that source. As of this catalog check (October 25, 2026), KimiSeek's live model catalog does not list Step 5 Preview or any StepFun model; this version is currently not available on KimiSeek. The OpenRouter ranking and availability mentioned in the report point to a third-party platform, not this platform. The Artificial Analysis intelligence index of 44, Elvis Saravia's capability comparison and pricing assessment cited in the report are all relayed content and have not been independently retested by this publication; model names, scores, and leaderboard versions may change with updates. The report does not give context length, API pricing, rate-limiting policy, or deployment hardware requirements; this information needs to be obtained from vendor documentation before integration.
Illustrations from the original report
