Capability-Based Model Router & Escalation

Nexum introduces a local-first capability router that optimizes latency, cost, and task success.

---

Capabilities & Tags

Models in your local Ollama library and cloud keys are tagged with capabilities:

• coding: High code synthesis accuracy (qwen2.5-coder, deepseek-coder).
• reasoning: Complex multi-step deduction and architecture (deepseek-r1, o1).
• vision: Image, screenshot, and diagram comprehension (llava, minicpm-v).
• quick: Sub-second local response for simple lookups (qwen2.5-coder:1.5b, llama3.2:3b).

---

Dynamic Self-Escalation (escalate_task)

When a fast local model encounters a problem beyond its parameter capacity (such as intricate multi-file architectural refactoring), it calls the escalate_task tool.

The router seamlessly migrates the turn to a cloud reasoning model with the full active transcript preserved.