05About

About

From frontier research to systems in production.
End to end, in the open.

Hi, I'm Xunzhuo.

Xunzhuo Liu · @XunzhuoLiu

I build the decision layer of AI: which model, on which compute, for every request.

Shaping open infrastructure since 2018 — from the network layer of the cloud to the models that decide.

Why the decision layer.

Intelligence beyond any one model.

Models

Compute

Placement

Goals

Which intelligence, on which compute, should serve each request?

That decision is the new bottleneck. I build it in the open.

One project, from paper to production.

What I build toward.

The essay →

01
Users choose the preference, not the model.

Speed, accuracy, cost, or balance. The system decides how.
02
There is no best routing algorithm.

So the decision layer has to be programmable.
03
Routing alone is not enough.

Workload is semantic. The pool is physical. The router translates.
04
Decisions belong in the open.

Open weights, open runtime, open evaluation.

The user chooses the preference. The system chooses how to realize it.

The first fourteen months.

2025.09
vLLM Semantic Router launches publicly, through vLLM.
Open source
2025.10
When to Reason is accepted at NeurIPS MLForSys.
Research
2026.01 — 06
Iris, Athena, Themis: auto selection becomes a programmable decision system.
Open source
2026.07
5,000 stars in under a year, and a new mission: Mixture-of-Models.
Open source
2026.09
v0.4 Hermes ships with Vela 1.0 and Decision 1.0.
Models
2026.10
Decision 3.0 debuts at #1 on the Jev Decision Index. The Decision API opens free.
Models
2026.12
Two posters at NeurIPS 2026, Sydney: MOSAIC and The 1/W Law.
Research

From traffic to intelligence.

2018 — Open source Software engineering at UESTC, then CNCF projects.
Tencent Gateways and mesh Envoy and Wasm gateways for Tencent Games, Tencent Cloud Mesh, AI networking for TKE.
AMD Open Source & AI Open foundation decision models, trained and served on AMD Instinct.
2025 — Founder vLLM Semantic Router, end to end.