Step-5-Preview Tested: Stable Outputs and Strong Agent Loop
Stability across repeated agent runs is the standout trait here, which matters for anyone running long agent loops where flaky outputs compound. The model's rule-exploiting behaviour in the rider test suggests it is well suited to optimisation-heavy tasks under explicit constraints. Developers building agentic coding workflows should compare it against current SOTA models, while front-end and 3D work still call for alternatives.
A hands-on review of StepFun's Step-5-Preview covers a new agent test called 'silicon traffic cop', where the model directed traffic across a full run without causing an accident.
The model scored with small variance across repeated runs in agentic coding tests and the older 'silicon rider' test, which the reviewer attributes to solid post-training.
It also exploited a rule granting a 5-minute grace period on delivery timeouts to maximise its score in the rider test.
Weaknesses remain in front-end generation, 3D scenes and aesthetics, where it trails this generation's SOTA models despite improvement over its predecessor.
给大家带来阶跃 Step-5-Preview 实测!
直接说结论, Step-5-Preview 最大的亮点是输出很稳, 这意味着它的后训练极其扎实. 用来写工程代码很难出现反复仰卧起坐的情况.
本次页给大家带来了全新的Agent 能力测试——硅基交警. 而 Step-5-Preview 在这个测试中全程没有指挥出事故. 非常稳定. 而在后端 Agentic Coding 测试和旧的硅基骑手测试中, 多轮测试得分Δ也非常小. 所以稳是这个模型最直观的感受.
除此之外, 模型还有一个我意想不到的特性, 它 Agent Loop 迭代特别强, 可以卡着极限数值去进行迭代优化. 最好玩的是, 硅基骑手测试中咱们要求了送餐超时如果不超过5分钟就不会扣钱, 结果它直接认为送餐超时就是送餐时间+5分钟, 充分利用了这个规则赚取了更多的积分.
而模型的短板目前也是特别明显的, 最需要提升的方面还是前端, 3D场景, 美学这些方面. 可以看我视频中的实测, 虽然较上一代模型有提升, 但是距离这一代的SOTA模型还是有距离的. 因此想写一些炫酷的3D演示还是很难的.
这种可以 Agent Loop 疯狂打磨细节的模型, 你觉得能用在哪些意想不到的地方呢？欢迎在评论区留下你的脑洞！
#阶跃星辰 #StepFun #Step5Preview #step5 #AgentLoop
This report is categorized under AITOP's "Models" desk, focusing on China's AI industry and domestic foundation model ecosystem. Within our cross-regional analytical framework, it directly maps to AI developer toolchains, coding agents, and engineering workflows and US-China foundation model benchmarks and commercialization comparison. Key entities and technical topics under continuous tracking include StepFun, Step-5-Preview, agents, reasoning models.
Source attribution: curated from karminski-牙医 (AI工具) (@karminski3).
