Routing

A serving system that decides which model, hardware tier, or execution path should handle a request so operators can balance cost, latency, safety, and specialization.

Opening summary

Routing means choosing where a request should go before the expensive work begins. A serving stack can send one request to a cheaper path, another to a faster path, and a third to a specialist path, all while presenting one product surface to the user.

At a glance

Released

June 2026

System type

System Routing

Related modules

What It Is

Routing is a serving-time control decision. The system inspects a request, applies simple policy or classification rules, and chooses which model, hardware tier, or execution path should answer it.

Where It Sits

It sits near the entry of the inference stack, between the incoming request and the concrete serving path that will execute the model. Some routing happens before model execution starts, while other routing can hand work from one serving path to another after a first pass reveals more about the request.

How It Works

A runtime usually begins with lightweight request features such as prompt length, account tier, safety risk, tool requirements, or whether the task looks general or specialized. A router then maps those signals onto an available path, such as a small fast model, a larger premium model, a specialist model, or stricter filtering and review stages.
Routing System Flow
Request and weight flow
Routing is a decision layer around serving paths. It chooses where work goes rather than changing the internal math of one model.

Practical Impact

Good routing lowers cost by keeping easy work off premium paths, lowers latency by avoiding oversized backends when they are not needed, improves safety by inserting stricter gates where risk is higher, and improves specialization by steering requests toward models or systems tuned for narrow jobs. This sits at a different layer than mixture-of-experts routing: serving-time request routing acts on whole requests or large serving stages owned by platform operators, while MoE routing acts inside one model and sends token representations to a few expert sublayers during the forward pass.
textrequestcostapproxtextcheap−pathsharecdotctextcheap+textspecialist−pathsharecdotctextspecialist+textroutingoverhead\\text{request cost} \\approx \\text{cheap-path share} \\cdot c_{\\text{cheap}} + \\text{specialist-path share} \\cdot c_{\\text{specialist}} + \\text{routing overhead}

Tags

References

  1. DeepSeek-AI. "DeepSeek-V4 Technical Report." 2026.