StepFun has released its flagship open-source model, Step 5 Preview, to users worldwide, according to a post on Alibaba Cloud’s official account cited by ChainCatcher. The model is positioned as a primary work model for real-world agent tasks and is described as being suited to AI coding, software engineering, professional knowledge work, and finance-related scenarios.
Alibaba Cloud said Step 5 Preview has 600B total parameters and 27B activated parameters. It uses a sparse mixture-of-experts, or MoE, architecture, supports a context window of up to 1M, and can process text, images, and video. The post also said the model recently ranked second globally on OpenRouter’s weekly and daily call-volume charts.
Alibaba Cloud says DAY0 adaptation was completed on launch day
On the day of the release, Alibaba Cloud announced that it had completed DAY0 adaptation for the model. The company said the setup was built on chip support, supernode instances, and joint inference optimization, with inference compute provided by Lingjun supernodes.
StepFun developed its own reinforcement learning training platform for agents, while the sandbox platform is managed by Alibaba Cloud. According to the post, the solution is based on Alibaba Cloud’s Agent Sandbox and the Kubernetes version of its container service. It is aimed at elastic compute for training workloads, can create more than 10,000 sandboxes in a single minute, supports up to million-level concurrency, is compatible with the E2B API, and offers Kubernetes-native interfaces.
Sandbox, storage, and security setup
The post said that through multi-level image caching, intelligent pre-scheduling, and optimization of the control chain, the system can deliver millisecond-level requests and create 30,000 sandboxes within minutes. Sandbox lifecycles are managed centrally by Alibaba Cloud’s backend.
On storage, each Pod starts with 15GB of free temporary storage. When usage exceeds 70%, another 30GB is added automatically each time, up to a maximum of 300GB.
On security, the default setup uses MicroVM-level isolation covering compute, network, storage, and permissions, while access to cloud products and large-model interfaces is handled under a zero-trust approach.
Performance figures cited in the post
Alibaba Cloud said image startup time was reduced by 90%, model evaluation cycles were shortened from weeks to days, and average EBS storage costs were cut by about 40%. With ACK One Fleet, the setup can support stable operation of sandboxes at a scale of hundreds of thousands at the same time, according to the post.