Ollama 0.40.17 is out, following 0.40.16. The release notes add Inkling 2 Mini from Thinking Machines, a 26B-parameter model that Ollama describes as having 5.4B active parameters, with looped layers, text and image support, tool calls, and thinking. It is listed as fitting on a Mac at about 20GB. In this release, inkling-2-mini is supported on Apple silicon only; PC support is promised in an upcoming release. Ollama on Apple silicon can now also serve multiple requests at once. A Vulkan change uses mmap for integrated GPUs, fixing “Failed to allocate pinned memory” load failures on shared-memory GPUs.
Other notes include launch-integration fixes for Claude Code deferred tool loading, OpenCode thinking options, Pi’s model list after sign-in, and a safer ChatGPT --restore path, plus a help print when ollama is run with no arguments in a non-interactive shell, and several sampler, Windows iGPU, and API schema fixes. I have not upgraded for this post. Ollama’s 5.4B “active” figure is the release’s wording; Thinking Machines’ own card splits that as 3.4B active and 5.4B effective with looping.
Review before you install: read the release notes yourself, back up your config, and try new versions somewhere you can afford to break first.
My take
The Apple silicon pieces are the ones I would look for first: Inkling 2 Mini, concurrent requests, and the Vulkan shared-memory fix if you have been hitting that load failure. Read the notes, then bump a copy you can afford to break. PC users will have to wait for Inkling.
God Bless, and Happy hacking!