Show HN: Model-literals, model-aliases, and preference-aligned routing for LLMs (docs.archgw.com)

Arch unveiled Arch-Router, a routing framework for multi-LLM deployments that offers three ways to route requests: direct model-based routing (explicit provider/model like openai/gpt-4o), alias-based routing (semantic names such as fast-model or reasoning-model that map to underlying models), and preference-aligned dynamic routing driven by a compact 1.5B Arch-Router model. The preference-aligned mode decouples route selection (human-readable policies expressed as Domain + Action, e.g., Domain: finance, Action: analyze_earnings_report) from model assignment, letting the router infer domain and action from prompt cues and pick the best model from your fleet according to configurable preferences. The system supports logging of routing metadata, load-based selection across mapped models, and mixing static aliases with dynamic routing. This matters because traditional routing often depends on brittle benchmarks and fixed pools; Arch-Router prioritizes human-preference alignment, transparency, and operational flexibility. Key technical implications: semantic inference uses task indicators and contextual cues to classify domain/action; routing policies are explicit and editable (no retraining to add new models); direct routing gives production predictability while aliases enable provider-agnostic experiments; dynamic routing improves cost, latency, and quality by matching intent to model strengths. The design targets low-latency, high-throughput production use cases and makes multi-model orchestration more controllable and auditable for real-world workflows.