New ADP KV cache connector release: GLM-5.3 and DeepSeek-V4-Flash on vLLM 0.28/0.29

A coding agent's context is accumulated work: the repository it has read, the tools it has called, and the decisions it has already made. Keeping that context reusable preserves GPU time for the next step or another user.

GLM-5.3 is built for complex coding and long-horizon tasks, with improvements over GLM-5.2 reported by Z.ai. It is an appealing choice for engineering assistants that carry a task through many tool calls. We recommend vLLM 0.28.0 for GLM-5.3.

DeepSeek-V4-Flash combines a mixture-of-experts design with a focus on efficient, long-context inference, making it an attractive option for document assistants and agents with substantial histories. We recommend vLLM 0.29.0 for DeepSeek-V4-Flash.

Both configurations have been verified on hardware. One connector code base supports these vLLM versions alongside the existing build, simplifying integration and maintenance. ADP keeps useful context outside GPU memory and makes it available for reuse when requests return to the same prefix.

The update includes test harnesses and image recipes alongside the connector changes. They give engineering teams a starting point for building the serving images and exercising the integration in their own environment.

This announcement reports model compatibility verified on hardware. Performance measurements in our previous connector release report used its stated models and test setups; they are not benchmarks for these new configurations.

We also fixed a number of issues and improved cache restoration and gateway connection recovery. These refinements make the connector more robust under concurrent workloads and help long-running inference services operate more reliably.