Starting a new AI model server, a cold start, can take minutes, which slows scaling up and recovery from failures.

An open-source checkpoint and restore system for Kubernetes that saves a fully started workload, including CPU and GPU memory, and restores it on another machine with the same GPU.

Asking a chatbot to return JSON for tasks like routing support tickets is slow and can break the expected format.

A guide and a new 'decision' pipeline for models that score a fixed list of allowed answers and return probabilities in one pass.

Cloud AI models are costly and raise privacy concerns, yet many simple assistant tasks don't need them.

A 0.8B-parameter model trained on top of Qwen3.5-0.8B for calling tools like search, calendar and email on the device.

Reasoning models for self-driving cars are too big and slow for the computer inside a vehicle.

Alpamayo 2 Edge, a 2B-parameter model that reasons only when needed, plus a judge model to score its reasoning.

Doctors labelling CT scans wait minutes for the AI to suggest outlines when it runs on a CPU.

A small port of the open-source MONAILabel server to AMD Instinct GPUs that works through PyTorch without new GPU code.