Reinventing Speed: Slim Installs, Thoughtful Models, and High-Leverage Infrastructure

Tools

vLLM sheds Ray dependency for a lighter install footprint

The vLLM team dropped v0.8.5 and the install process is now 50% lighter on disk since they cut a big Ray dependency. If you are running inference pipelines in production, update now; your Docker image size and install time will thank you.

SmolLM3 lands in llama.cpp for universal hardware access

Llama.cpp has merged support for SmolLM3, opening the door to running this high-performance dataset-driven model on basically any hardware. For builders obsessed with edge performance, this is the most efficient path to getting a modern 3B model running on a laptop right now.

llama.cpp gets speculative decoding for real-time VLMs

Running vision models locally used to be a pain, but this patch makes llama.cpp fast enough for real-time video feeds. This is the missing bridge for practical multimodal agents that need to 'see' the world without sending every frame to the cloud.

Model

Microsoft's Phi-4-reasoning-plus brings 'thinking' to small models

Microsoft proves that explicit reasoning traces work for tiny models too, not just massive ones like R1. This is a huge deal for edge deployment; if you can run a 3.8B model that actually 'thinks' through problems, you don't need a 70B powerhouse for complex tasks.