Open Source AI··
BlueskyNewsRedditNews
Open Source AI's Quiet Infrastructure Moment
Practitioners running frontier-scale open models locally are making closed alternatives structurally redundant for a growing segment of professional developers.
The open versus closed AI debate has been conducted for years as a question of values: who controls weights, who can audit training data, who bears liability. That framing has given way to a more immediate question — whether local hardware can run frontier-scale models at production-viable speeds. The answer, as of this week's practitioner benchmarks, is yes for a meaningful segment of professional workloads. A paged mixture-of-experts engine running Qwen3.5's 397-billion-parameter model at nearly 3 tokens per second on a 64 GB M1 Ultra [3] is not a demonstration…