Chinese
Yushan AI 正式使用 NVIDIA H100 GPU 開始訓練 3B 主權 AI 模型。這次訓練的重點， 不只是把模型跑起來，而是把資料、程式、訓練流程與 checkpoint 管理串成一套可以持續迭代的工程基礎。
從 Jupyter workspace 中的 yushan-core.zip，到 yushan-3b-instruct-v0.1 的 checkpoint 持續產出，可以看到模型訓練已經進入實際執行階段。adapter_model.safetensors、 tokenizer.json、optimizer.pt、scheduler.pt 與 trainer_state.json 這些檔案，代表訓練狀態、 參數更新與後續微調流程都開始被系統化保存。
主權 AI 的意義，不只是擁有一個模型名稱，而是能夠掌握訓練資料、模型架構、推論部署與在地應用場景。 對台灣而言，3B 模型是務實的起點：規模足以承載許多語言與產業任務，也更適合在成本、速度與可部署性之間取得平衡。
這一步代表 Yushan AI 從概念、資料整理與工程準備，走向真正的模型訓練循環。接下來更重要的工作， 是持續評估模型表現、強化繁體中文與台灣語境能力，並把模型能力導入醫療、企業、邊緣運算與產業應用。
English
Yushan AI has started training its 3B sovereign AI model with NVIDIA H100 GPUs. The milestone is not only that the model is running, but that the data, code, training workflow, and checkpoint management are becoming part of a repeatable engineering foundation.
From the yushan-core.zip package in the Jupyter workspace to the generated yushan-3b-instruct-v0.1 checkpoints, the model has entered the execution stage. Files such as adapter_model.safetensors, tokenizer.json, optimizer.pt, scheduler.pt, and trainer_state.json show that training state, parameter updates, and future fine-tuning workflows are being preserved systematically.
Sovereign AI is not just about naming a model. It is about control over training data, model architecture, inference deployment, and local application contexts. For Taiwan, a 3B model is a pragmatic starting point: large enough for many language and industry tasks, while still balanced across cost, speed, and deployability.
KEY SIGNALS
- NVIDIA H100 GPU training is underway
- Yushan 3B instruct checkpoints are being generated
- Training state, tokenizer, adapter, and optimizer files are preserved
- The work advances a practical sovereign AI foundation for Taiwan