The Machine Learning ecosystem devised general purpose file formats and runtimes to share compute graphs and weights, and solve deployment issues since the 2010s. Now hardware or domain specific runtimes are being added to provide improved capabilities.

Qwen 3.8 27B is an exciting release, but still too large for common dev machines. But I can find interesting models to run locally.

The landscape of coding agents tooling is fragmented, and there are no definitive winners.

Testing local LLM inference speed on consumer hardware — prefill, decode, and KV cache impact across three devices and two models.

Following a previous mistake, I find subqueries can beat CTE performance, but are still worse for the disk