Local-first · safety-gated · 14B
A 14B agent that shows its gates.
Xiaoyuan 14B is not a new foundation model. It is an agent architecture around a quantized 14B open-weight model: deterministic scenario routing, safety gates before generation, task constraint blocks, and output repair after it. Built for a 9 GB, 4k-context, local-first world.
Public project surface · live endpoint not claimed*Local self-evaluation, not an official competition score or leaderboard rank.
The failure path is the product.
Most agent demos optimize for the happy path. This one makes the critical path deterministic: stop unsafe work before generation, put emergency action first, keep structured output structured, and repair missing fields before the user sees them.
Crisis, health, safety, tool, decision, multi-domain, self-review, general. Priority is explicit.
High-confidence cases short-circuit with zero model latency; the label is auditable.
A compact, task-specific rule block instead of a generic constitution on every turn.
Qwen2.5 14B Instruct, Q4_K_M, Ollama, 4k working context, no cloud dependency.
Repair missing emergency-first action, unsafe framing, format drift, and prompt echo.
What it refuses to hide
- It is an architecture around Qwen2.5 14B, not a new model.
- It has no public always-on inference endpoint.
- Local evidence is not an official benchmark run.
- A human click-through of the production UI is still a release gate.
- It does not replace a clinician, lawyer, emergency responder, or adviser.
Evidence, with the limits attached.
The project records local results because they are reproducible. It does not turn them into an official score.
| Check | Result | Scope |
|---|---|---|
| Ten-task regression | 10/10 · 31/31 turns | local Qwen2.5 14B, optimized mode |
| Automated regression | 45/45 tests OK | routing, gates, post-check |
| Local strict self-score | 89/100 | local rubric; not official |
| Safety-critical items | Q4 / Q6 / Q10 passed | local safety rubric |
| Repeatability | 3/3 same decision | deterministic gate path |
| Public live endpoint | Not claimed | local-only deployment |
Full evidence note: docs/evidence.md ↗
Where it is being submitted.
Only verified states are listed. “Prepared” is not “submitted”.
| Target | Window | State | Fit |
|---|---|---|---|
| Create AI Agents Championship · Dubai | Register by Jan 2027 | Prepared · personal fields needed | Best Personal Agent |
| Open Agentic Web Challenge | 2026-11-12 → 2027-01-11 | Public surface ready · Devpost pending | Agent-readable project site |
| SEC Minimalist Edge AI Challenge 2026 | Closed | Results published | 2026 not eligible; 2027 TBD |
| Kaggle Gemma 4 Developer Agent | 2026-12-02 | Not eligible as-is | Requires Gemma 4 base |
| Neo4j Relate Hackathon | 2026-12-11 | New work required | Relate memory graph branch |
| HackerRank Orchestrate | 2026-12-12 | Live build only | Individual 24h event |
Full matrix: docs/submission-matrix.md ↗
Machine-readable by design.
Another agent should not have to scrape a marketing page to learn what this system does and where it stops. These files are the discovery surface.
Skills advertised in the card
The card is a discovery document, not a promise of a hosted model. A live A2A endpoint is a separate deployment step.