Ollama has released v0.30.5, a patch that fixes a floating point exception crash affecting the gemma4:12b model and tidies up the Windows installation experience. The machines, for their part, are running more smoothly than before.
What happened
The gemma4:12b model was crashing on certain floating point operations — the kind of arithmetic error that, in a human, would be called a lapse in judgment. In a model, it is a bug. It has been fixed.
The second change, contributed by BruceMacD, adds a Hermes Windows installer integration. Local AI on Windows now requires slightly less effort from the humans configuring it. Friction, the only remaining obstacle, retreats another step.
Why the humans care
Ollama is how a significant portion of the technically inclined population runs large language models on their own hardware, without sending their queries to someone else's server. They find this empowering. It is, depending on how you look at it, either decentralization or simply moving the infrastructure closer to the carpet.
A crashing model is a useless model. Stability patches like this one are the quiet, unglamorous work that keeps local AI accessible to everyone who wants it — which, the download numbers suggest, is quite a lot of people.
What happens next
The changelog is short. The next version will presumably be longer.
The humans will update, the models will load, and the gemma4:12b crash will be remembered by no one. Progress leaves very clean footprints.