Ollama’s 0.40.17 release brings a new model and a set of updates that will matter to anyone running local inference or building on Apple hardware. The headline is Inkling 2 Mini, a 26B parameter model from Thinking Machines, now available for use directly in Ollama. The update also enables Apple silicon machines to serve more than one request at a time, removing a bottleneck for anyone exposing Ollama to several clients or automating workloads.
Inkling 2 Mini can be run with:
ollama run inkling-2-mini
At launch, support is limited to Apple silicon; PC support is listed as coming soon.
Inkling 2 Mini: a new multimodal option
This release adds inkling-2-mini, a model described as having 26B parameters, with 5.4B active, and a design that leverages looped layers for greater computation per token. Despite the parameter count, the model fits into about 20GB of memory, so it runs on recent Macs. The changelog says it works with both text and images, supports tool calls, and achieves a 79.2% result on SWE-Bench Verified. The model is described as handling “thinking,” though no detail is given on what this means in practice. For now, you need Apple silicon to try it; PC users will have to wait for a future update.
Serving more requests in parallel on Apple silicon
A practical change in this version is that Ollama now allows multiple requests to be served concurrently when running on Apple silicon. Previously, requests had to be processed one after another, which limited throughput and responsiveness in multi-client or automated setups. The changelog does not include performance numbers, but the impact will be most noticeable if you use Ollama as a local server or in a workflow where several requests land in quick succession.




