Starting a new AI model server, or recovering a failed one, can take minutes while weights load and code compiles.

An open-source Kubernetes tool that saves a fully started workload's CPU and GPU memory and restores it on another machine with the same GPU and driver.

Very long inputs strain compute, memory and links between GPUs at once, so tuning one piece is not enough.

A set of changes across the serving stack on eight GPUs, including splitting tokens across GPUs, a smaller 8-bit memory cache and lighter data exchange between experts.

Cloud AI is costly, can change without notice and sends user data away, while many simple tasks don't need a big model.

A small model trained on top of Qwen3.5-0.8B to call tools for tasks like search, calendar and email, running on-device.

Testing code against real AWS services means deploying and waiting, which slows developers and coding agents.

A local emulator written in Rust that answers the same AWS APIs over a local HTTPS endpoint, with an in-process mode for Rust tests.