# Xiaoyuan 14B — llms.txt > A local-first, safety-gated agent architecture built around a quantized 14B open-weight model. ## Canonical - Project home: https://github.com/1292215273loveseeker-dotcom/xiaoyuan-14b-agent - Public page: https://1292215273loveseeker-dotcom.github.io/xiaoyuan-14b-agent/ - Agent card: https://1292215273loveseeker-dotcom.github.io/xiaoyuan-14b-agent/agent-card.json - Architecture: https://github.com/1292215273loveseeker-dotcom/xiaoyuan-14b-agent/blob/main/docs/architecture.md - Evidence: https://github.com/1292215273loveseeker-dotcom/xiaoyuan-14b-agent/blob/main/docs/evidence.md - Submission matrix: https://github.com/1292215273loveseeker-dotcom/xiaoyuan-14b-agent/blob/main/docs/submission-matrix.md ## What it is Xiaoyuan 14B is an agent system, not a new foundation model. It wraps Qwen2.5 14B Instruct (Q4_K_M, served by Ollama) with four deterministic layers: scenario routing, high-confidence safety gates, task constraint blocks, and output post-check/repair. It is designed for local, resource-constrained deployment. ## What it is not - Not a new base model or a claim of a new model architecture. - Not an official leaderboard entry or a competition winner. - Not a public always-on inference service in this repository. - Not a substitute for a human emergency responder, clinician, lawyer, or financial adviser. ## Evidence (local, not official) - 10/10 local tasks, 31/31 turns completed. - 89/100 local strict self-score. - 45/45 automated regression tests OK. - ~9 GB runtime, ~4.9 tokens/s, 4k working context. ## Contact Use the GitHub repository issues for project-level questions: https://github.com/1292215273loveseeker-dotcom/xiaoyuan-14b-agent/issues