Alibaba Cloud helps Jielue Xingchen build a computing power factory, creating 30,000 sandboxes in minutes

The official account of Alibaba Cloud stated that the large model company Step Star has globally released its flagship open-source model Step 5 Preview, positioned as a primary working model for real-world intelligent agent tasks, excelling in AI programming, software engineering, professional knowledge work, and financial scenarios. The model has a total of 600 billion parameters and 27 billion active parameters, utilizing sparse MoE, with a maximum context window of 1 million, capable of processing text, images, and videos. The article mentions that it recently ranked second globally in both weekly and daily call volume on OpenRouter. On the day of the release, Alibaba Cloud announced the completion of DAY0 adaptation, optimized for joint inference based on chips, super node instances, and models, with inference computing power provided by the Lingjun super node.

Step Star has developed its own intelligent agent reinforcement learning training platform, with the sandbox platform managed by Alibaba Cloud. The solution is based on Alibaba Cloud's Agent Sandbox and Kubernetes container service, aimed at elastic computing power in training mode, claiming to create tens of thousands of sandboxes in a single minute, with a maximum concurrency of up to one million, compatible with E2B API, and providing native Kubernetes interfaces. Through multi-level image caching, intelligent pre-scheduling, and control link optimization, it achieves millisecond-level application and minute-level creation of 30,000 sandboxes, with the sandbox lifecycle uniformly managed by Alibaba Cloud's backend.

In terms of storage, each Pod starts with 15GB of free temporary storage, automatically increasing by 30GB each time when usage exceeds 70%, up to a maximum of 300GB. For security, MicroVM-level isolation is adopted by default, covering computing, networking, storage, and permissions, with zero-trust access to cloud products and large model interfaces. The results provided in the article include: a 90% reduction in image startup time, model evaluation cycles compressed from weekly to daily, and an average savings of about 40% on EBS storage costs; relying on ACK One Fleet, it can support the stable operation of hundreds of thousands of sandboxes simultaneously.