Ollama 0.40.6 stops leaving runners behind on Windows and opens Q2_0 model files

Ollama 0.40.6 is a small release with two fixes that matter to two different groups of people. If you run Ollama on Windows by typing ollama servein a terminal, it no longer leaves its runner processes behind when it stops. And if you have model files in the very compact Q2_0 format, Ollama can now read them. The release page carries the date 16 October, 22:00 , which is when the page opened as a pre-release; the installers it offers were uploaded on 17 October at 00:57 UTC , and its changelog line still runs to the release candidate, v0.40.6-rc0. Below is what each fix does, how to tell whether it concerns you, and what to do.

What changed

Windows: no more runners left behind. When Ollama answers a request, the model does not run inside the main ollama program but in a separate helper process, a "runner" — llama-server or an MLX process. Ollama starts these runners and is meant to end them when it stops. From this release, Ollama places its runners in a Windows Job Object configured with JOB_OBJECT_LIMIT_KILL_ON_JOB_CLOSE: a group of processes that Windows itself closes when the program that owns the group ends. In practice, Windows now ends thellama-serverand MLX processes Ollama owns when Ollama exits, even when Ollama is forced to stop and gets no chance to tidy up. The developers also made the fix try again on every model load, so a passing failure on the Windows side does not leave it stuck.

Q2_0 model files now open.GGUF is the file format Ollama and llama.cpp use for models, and inside it each layer's weights are stored in a "quantization" type — a way of squeezing the numbers into fewer bits. Q2_0 is one of the smallest: it stores 64 weights in 18 bytes , a little over two bits per weight. The llama.cpp build already inside Ollama 0.40.5, b11669, could run such files, but Ollama's own GGUF reader did not recognise the type. This release adds it. The change names one family it is needed for: Prism ML's Bonsai 2 models, in 2B, 4B, 9B and 27B sizes, whose Q2_0 files are the ones upstream llama.cpp can run. On Apple silicon, all four sizes were created and run with thinking, vision and tools, while 0.40.5 fails on the same files.

The rest of the list is small:

• A change titled "Launch clef in Ollama's cloud" is a documentation update, made because Clef is now available on Ollama's cloud. The cloud tags themselves came in the previous release; our itemOllama 0.40.5 brings NVIDIA's Nemotron 3.5 Super and its smaller sibling, Bolt covers them.
• The --askmode now uses identifiers for its question keys. That is the mode that puts questions to decision models from the terminal, explained in our itemOllama 0.40.4 lets you ask decision models questions straight from the terminal. The page says nothing more about it, and nothing on it tells you to change the commands you type.

What it was before

The changelog line runs from v0.40.5. Before this release, the two problems looked like this.

On Windows, the desktop app was never affected: it shuts the Ollama server down gracefully and gives it a chance to clean up its child processes. The trouble was for people runningollama serveat the command line, who, the change says, lack good command-line tools on Windows to end those processes cleanly. When such a server was stopped abruptly, its runners could stay behind on their own — the change's title calls them orphaned.

With Q2_0, the failure was blunt: ollama createand model loads stopped with an error saying an unsupported tensor size overflows. The file was fine; the reader simply did not know type 42, or file type 41.

What it means for you

If you run ollama serve on Windows from a terminal, a script or a scheduled task, update. A runner left behind keeps running with the model it had loaded, and a model can take gigabytes of your graphics card's memory, so the next model you load may find less room than you expect, or run slower because part of it no longer fits. Before you update, open Task Manager, go to the Details tab, and look for llama-serverprocesses while Ollama is not running. If you find any, end them there, then install 0.40.6. From now on, Windows ends them for you when Ollama exits, however it exits. If you only ever use the Ollama desktop app, this fix changes nothing for you.

If ollama create refused a Q2_0 file with the "size overflows" error, update and try again.The same file should now load. Q2_0's appeal is size: at a little over two bits per weight , a model takes far less memory than the same model in a common four-bit or eight-bit type, which can be the difference between a model fitting on your machine or not. The cost is that squeezing a model this hard usually loses some quality, so try it on your own tasks before relying on it. Our guideWhich Machine Runs a Model Locallyexplains how a file's size decides whether it fits on your machine. The tests named in the change were run on Apple silicon ; the change does not report results on Windows or Linux machines.

If neither applies to you, there is no urgency. The other two entries are a documentation update and a small change to --ask.

The source

The release page on GitHub, Ollama v0.40.6, and its list of downloadable files with their upload times; the changes behind the two fixes, pull request #18500 (Windows runners) and pull request #18967 (Q2_0); and the documentation change, pull request #19036.