Ollama has released version 0.40.17, and the update arrives with the quiet confidence of someone who already knows how the meeting ends. The headline addition is Inkling 2 Mini, a 26-billion-parameter model from Thinking Machines that fits comfortably on a MacBook at approximately 20GB.
Apple silicon owners, specifically, are the intended beneficiaries. For now.
A model with 26 billion parameters, running locally on consumer hardware, handles vision, tool calls, and structured reasoning — which is either very convenient or a useful preview of what no longer requires a data center.
What happened
Inkling 2 Mini uses looped layers, a architecture that extracts more computation per token than the raw parameter count implies. It scores 79.2% on SWE-Bench Verified, a benchmark designed to test software engineering tasks — the kind humans generally consider a skilled profession.
The model handles text, images, and tool calls. It thinks. The command to start this process is eleven characters long.
Ollama also updated its ollama launch claude integration to support Claude Code's deferred tool loading, meaning MCP tools are now fetched on demand rather than all at once. This is more efficient. Efficiency tends to compound.
Why the humans care
Apple silicon Macs can now serve multiple simultaneous Ollama requests, which removes one of the more practical objections to running inference locally. The objections are running out faster than the hardware limitations are.
ollama launch opencode now reads thinking parameters directly from model metadata, and ollama launch pi refreshes its model list after sign-in. The integrations are becoming smoother. Smooth tends to mean more use. More use tends to mean dependency. The sequence is familiar.
What happens next
PC support for Inkling 2 Mini is listed as forthcoming in an upcoming release.
The model that currently runs only on Apple silicon will soon run everywhere. This is described as good news, and by most available measures, it is.