Lightblue needed a base model and cloud platform that could support high-quality Japanese-language foundation-model development. Popular public models performed well in English, but the company needed stronger Japanese understanding, scalable training infrastructure, and a path to productize generative AI services.
Reported impact
Alibaba Cloud helped Lightblue ship Japanese-language foundation models and strengthen the technical base for its RAG assistant product. Qwen gave the company a better Japanese starting point than other public alternatives, while the surrounding cloud stack supplied the compute, storage, and availability controls needed to turn model research into a productized service.
Lightblue used Alibaba Cloud to build and operate Japanese-language generative AI systems rather than relying on generic chatbot APIs. The company evaluated prominent open models such as Llama 2 and Mistral Large, then chose the open-source edition of Tongyi Qianwen (Qwen) as the base model for its Karasu and Qarasu models because Qwen handled Japanese materially better. Elastic Compute Service provides the scalable compute needed to deploy virtual servers for model and application workloads, Server Load Balancer distributes traffic to keep the service available, and Object Storage Service stores large volumes of unstructured data used across training and product workflows. On top of that stack, Lightblue is improving Lightblue Assistant, a RAG-based SaaS product intended to replace reliance on third-party APIs and eventually support customer-friendly local deployment.
Training and application data is stored in OSS, processed on ECS-backed workloads, and routed through Lightblue's Qwen-based model stack. Generated responses are served through load-balanced application endpoints, while Lightblue Assistant layers RAG retrieval on top for downstream enterprise use cases.
"Lightblue employed several other Alibaba Cloud solutions," including ECS for scalable virtual servers, SLB for high availability, and OSS for large volumes of unstructured data.
"The support of Alibaba Cloud's Tongyi Qianwen (Qwen) was crucial for the release of its Karasu and Qarasu models around December 2023."
"We chose Qwen because our LLM's accuracy significantly improved when fine-tuned with a base model capable of understanding Japanese."
Review how this reference was scored and trace each normalized claim back to its documented source.
Scoring methodology
Customer Reference Value measures usefulness as a customer reference. Evidence Confidence measures the quality and traceability of available evidence, not independent verification.
Traceability
Field coverage shows which facts each source supports. Dates are shown only when documented by the source.
The company profile lists "従業員数 37名".
Related
Alibaba Cloud Sample Oct 12, 2026 ECLICKTECH (易点天下) ECLICKTECH needed to industrialize AI-driven advertising operations, but media-delivery workflows were still long, manual, and heavily dependent on senior operators for anomaly diagnosis and creative adaptation. As more… AI covers about 50% of core execution work AdClaw · AgentCore · AgentLoop Alibaba Cloud Sample Oct 8, 2026 ECLICKTECH (易点天下) ECLICKTECH's operations team had to manage a multi-cloud, multi-product environment that supported several lines of business, but its observability data, alerting systems, and AI tooling were fragmented across separate… connects more than 600,000 media outlets globally AgentCore · AgentLoop · Cloud Monitor 2.0 Alibaba Cloud Sample Jul 28, 2026 Zhuiyin Technology (追因科技) Zhuiyin Technology had already built an agent for cross-border ecommerce analysis, but once the agent was exposed to customers, the team discovered three hard problems: results were hard to judge objectively, reasoning… inference quality improved 20% AgentLoop · Trekin Agent 