Katanemo releases Arch‑Router 1.5B to streamline multi‑LLM orchestration

Katanemo Labs has released Arch‑Router 1.5B, an open-source routing model designed to select the optimal LLM for a given query. Unlike retraining-intensive alternatives, Arch‑Router achieves high routing accuracy out of the box by aligning user input to developer-defined “domain→action” policies. The model hits 93.17% accuracy in preference-aligned routing, outperforming GPT-4 and Claude 3 Opus by over 7 percentage points in benchmark tests. It’s available now on Hugging Face under a research license.

Routing in multi-agent AI systems has typically relied on brittle heuristics or costly fine-tuning. Arch‑Router reframes the problem: It uses an instruction-tuned LLM to learn the mapping between a user's real-world intent and a developer's routing rules. This allows builders to create dynamic, modular LLM systems that can pivot between tools or APIs based on nuanced context, without modifying the router itself. With just 1.5 billion parameters, it’s also small enough to run efficiently in production.

The release positions Katanemo as a pragmatic player in the infrastructure layer of the AI stack. While incumbents focus on massive foundation models, Arch‑Router tackles a concrete developer pain point. Its strongest use cases are in retrieval-augmented generation (RAG), autonomous agents, and low-latency AI routing in SaaS applications. By open-sourcing the model and publishing benchmark results, Katanemo is making a direct appeal to the AI dev community to adopt and extend the tool.

Imagine your company support chatbot that needs to answer a customer’s billing question, schedule a delivery, and troubleshoot a product. Each of those tasks may require a different AI tool behind the scenes. Without smart routing, the system might waste time sending every query to the same large model or break when rules get too complex. Arch‑Router solves this by acting like a traffic cop. It looks at the user’s request and decides, in real time, which AI tool is best for the job based on clear, human-written rules. No need to rewrite the whole system or retrain anything. It just works. It’s faster and smarter.