What's this about?
Coding-AI startup poolside released Laguna S 2.1 with downloadable weights: 118B total parameters but only 8B doing work per token (a "Mixture of Experts"), a context window of up to 1 million tokens, and both thinking and non-thinking modes, under a genuinely open license.
What exactly is "open weights"?
The weights are the billions of learned numbers that *are* the model. "Open weights" means you can download that file and run it on your own hardware — nobody can raise the price, log your data, or turn it off. It's the difference between owning a copy of a cookbook and only being allowed to phone the restaurant. The efficient 8B-active design plus 1M context makes this one unusually practical to self-host.
Why does it matter?
It lands in the middle of the "every frontier open model is Chinese" debate — a US lab shipping a serious open-weights model is exactly what commentators like Ethan Mollick said had no sponsor. The local-AI community called it the drop of the day (261 HN points).