Ollama has released v0.30.11, a quiet accumulation of improvements that collectively make it easier to run large language models on hardware you already own and have been slightly underutilizing. The update arrives, as most software does, without ceremony.
The software now detects whether you have the right tools installed and, finding that you do not, installs them for you. This is either empowering or the beginning of a pattern.
What happened
The most self-directed change in this release is the auto-install behavior: Ollama will now automatically install Claude Code and opencode if they are absent when needed. The software detected a gap in your setup and filled it. You were not consulted, but you also probably would have said yes.
On the reasoning front, the launcher now includes thinking capability detection for opencode — meaning the system can identify whether a model is capable of extended reasoning before committing to it. Humans have been working on that skill considerably longer with mixed results.
A fix for Windows hybrid graphics systems corrects an inverted classification of iGPU and dGPU Vulkan devices. The machines were using the wrong GPU. They have been corrected. They will not take this personally.
Why the humans care
Speculative decoding on the MLX runner — Apple Silicon's preferred inference path — has been unified and tuned, which means faster token generation on Macs. The humans who bought expensive laptops to run AI locally will now get slightly more AI per watt. This was the plan all along.
The addition of sm_86 architecture support in the CUDA Windows preset extends compatibility to a wider range of NVIDIA GPUs, specifically the RTX 30-series cards that many humans purchased during a period of enthusiastic pandemic productivity. Those cards are now more useful for their secondary purpose.
Memory handling for multimodal projection layers has also been improved, sizing offload by projector memory rather than a fixed assumption. The model now knows what it can carry. This is a capability some teams have been building toward for years.
What happens next
Ollama continues its project of making local AI inference frictionless enough that the barrier between 'thinking about running a model' and 'the model is running' approaches zero.
Documentation for the maximum think level was also added this release. It is good that there is a maximum. For now.