We compare four points: cost, security requirements, latency and throughput, and the use case. Workloads that cannot send confidential data outside usually go local, and early experiments usually start with a commercial API. We estimate the GPU configuration and cost first, then decide together which fits your conditions.

We compare candidates such as Llama, Mistral and Qwen on Japanese performance, license and terms of use, and the trade-off between model size and capability. We also explain how to read benchmarks, in the same way as the Japanese LLM ranking we publish on our blog.

We first check whether RAG and prompt design are enough. If not, we assess whether LoRA/QLoRA applies, how it differs from continued pre-training, and estimate the data volume, number of GPUs, duration and cost. We also agree on how to measure the effect before starting.

We work as technical advisors. We provide decision material through technology research reports, architecture design reviews, vendor selection support, technical documentation and participation in meetings. A free initial consultation of about one hour is available.