Public artifacts

Hatsunama AI Research

Hatsunama conducts independent artificial intelligence research and development spanning AI models, AI agents, local AI, agent harnesses, tool use, computer-use systems, AI evaluation, and applied AI engineering.

Models and data

• Qwen3 tool-calling LoRA
• Qwen3 repaired checkpoint
• Caption Studio dataset
• Caption Studio Hy-MT2-1.8B Q8 candidate — an experimental local subtitle-translation artifact derived from Tencent Hy-MT2-1.8B. The public model card reports mechanical checks; meaning was not independently graded, and production app integration or final APK/AAB acceptance is not claimed. Read the evidence and limits.

These are distinct artifacts. Follow each model card for its provenance, limitations, and intended use.

Evidence behind the work

• Qwen3 tool-calling LoRA: the model card reports 15/16 hold-out cases, identifies a remaining unbound-tool call, and labels its BFCL-style scoring as a simplified check. These are publisher-reported results, not an independent replication or a task-completion benchmark.
• Qwen3 embedding repair: the card documents replacing non-finite embeddings using the upstream Qwen model while preserving the modified transformer blocks. This is a modification of existing models, not a new foundation model.
• Caption Studio and Caption Action: public source repositories for local captioning and live caption/translation software.

Read the research notes for the engineering questions and limitations behind these artifacts.

Software

Current free-app research — Hatsunama’s self-taught work on captioning, translation and local text-search projects, with development status and evidence limits.

• Caption Studio — local captioning app
• Caption Action — live caption and translation app

Read the research approach or about Hatsunama.