Day-Zero Model Support

Day-zero model support is the inference-infrastructure requirement that a newly released model should be usable by customers on the first day it appears. In E247｜对话盛颖：xAI，Infra的浪漫，SGLang，开源，平权与“甄嬛传”, [[ShengYing|盛颖]] uses this to explain why SGLang and [[RadixARC|Redix ARK]] have to track new architectures closely rather than optimizing only for stable existing models.

The source’s example is DeepSeek V4, where architectural novelty required significant adaptation and rewriting. The concept therefore connects release-day user expectations with Model-Infra Co-Design, Inference Acceleration Stack, and the maintenance burden inside Open Source AI Infrastructure.

Key Claims

• Users increasingly expect new models to be served immediately rather than after a long engine-support lag.
• Model architecture changes can break assumptions about cache, scheduling, kernels, memory layout, and rollout behavior.
• Day-zero support turns inference infrastructure into a live compatibility product, not a one-time performance layer.
• Open-source engines need maintainer capacity and production users to keep pace with frontier-model churn.

Connections

• SGLang, [[ShengYing|盛颖 / Sheng Ying]], and [[RadixARC|Redix ARK]] - source case.
• Inference Acceleration Stack, Model-Infra Co-Design, and AI Infrastructure As Product - infrastructure adaptation frame.
• Prefix Caching, Radix Attention, and Agent Inference Workload - serving assumptions that new architectures can stress.
• DeepSeek and Open Source AI Models - model ecosystem pressure behind compatibility.