Ollama 0.40.4 lets you ask decision models questions straight from the terminal

Ollama 0.40.4 brings decision models to the command line: you can now use them directly from the terminal with ollama run, to ask a yes/no question, sort a piece of text into labels, or score it against a rubric. Until now you reached these models through an API request; now one line in a terminal is enough. The release page carries the date 12 October, 19:00 , but that is the day the page was opened, while the release was still a pre-release; the installers an update downloads were uploaded on 14 October at 00:52 UTC. Below is what the new command does, what else changed, and who should update.

What changed

A decision model answers with a choice, a probability or a score rather than with sentences; our item Ollama 0.35 adds decision models that answer with choices, probabilities and scores instead of text explains them and the API they came with. What 0.40.4 adds is a way to use them without writing that request. The page's example runs the model nimble with the option --ask and the question "Is a refund being requested?", and feeds it the text of a support ticket from a file named ticket.txt. The<in that line is the shell's way of saying "read the input from this file", so the same line works with any text file you have. The three kinds of question the page names are the ones the API already had: a yes/no question, a choice among labels, and a score against a rubric — a rubric being a list of criteria you want the text judged on.

Two changes underneath support this. Ollama now ships a version of llama.cpp — the engine that runs models on most machines — with native support for decision models , and on an Apple Silicon Mac, Kev decision models now run on MLX, the engine Apple's chips use.

The rest of the list is shorter work:

• ollama launch, the command that starts a coding tool wired to a model, can now be used with usage credits without upgrading to a paid plan. The page does not say how credits are obtained or what they cost.
• ollama launch codex now shows the models installed on your machine in its model picker, and /modelinside the Codex CLI switches between them.
• Inference — the model producing its answer — is faster because Ollama reuses manifests it has already checked and models it has already loaded. A manifest is the small file that lists the parts a model is made of.
• Multimodal embeddings now work through llama-server , and multimodal embedding models with small context windows load again. An embedding model turns content into a list of numbers that captures its meaning; a multimodal one accepts images as well as text.
• On NVIDIA graphics cards, the MLX runner reads prompts to Gemma 4 faster — that reading stage, before the first word of the answer, is called prefill.
• The MLX runner clips long Laya states instead of rejecting them , and speculative decoding there is more reliable, including the handling of recurrent state after a rollback. Speculative decoding is a speed-up in which a quick guess at the next few words is checked by the full model and rolled back where it was wrong.
• The README and documentation now describe privacy and the cloud services in clearer words.

What it was before

The page lists its changes as those since v0.40.3. Before this release, a decision model was something you called from code, through the endpoint described in our 0.35 item; the terminal route is what this release adds. The 0.40.4 page says nothing about the background model upgrade that 0.40.3 paused, so as far as the page tells you, that pause still holds. Our itemOllama 0.40.3 pauses the background model upgrade that 0.40.2 brought in explains the pause.

What it means for you

If you have wanted to try a decision model but did not want to write an API request first, update and start with the page's own line: put a short text in a file, pull nimbleif you do not have it, and ask it a yes/no question about the file. It is the quickest way to see what these models are good at — a fast, plain answer to a narrow question — before you build anything on them.

If you sort or screen text in a script — tickets, messages, comments — the terminal command means a shell script can now call a decision model one file at a time. Keep each question narrow and each label clearly distinct from the others, so the answer means one thing.

If you code with Codex through ollama launch codex, update to get your local models in the picker. If you useollama launchwithout a paid plan, check whether usage credits change what you can run ; the release page does not give the terms.

If you run embedding models with images, or Gemma 4 through the MLX runner on an NVIDIA card, this release fixes or speeds up your case directly. For everyone else, the faster reuse of loaded models is a quiet gain that needs nothing from you.

The source

The release page on GitHub, Ollama v0.40.4, and its list of downloadable files with their upload times.