Ollama’s 0.40.11 release notes, dated 24 October 2026, introduce d3: models that select categories or score options rather than write a conversational reply. For a small support team, this is a reason to test one narrow step: deciding who should receive an enquiry. It is not a ready-made support system or evidence that automated routing already works for your business.
What changed
Ollama presents vLLM Semantic Router’s d3 family in 2B, 4B, 9B and 27B sizes. A model receives input and the options you define, returning scores and probabilities. The official d3-mini model card explains the distinction from a chatbot: the output is a choice, a yes/no answer or a scale rating, without generating a text response.
The release-note date does not mean the model itself originated that day: the package page specifies compatibility with Ollama 0.40.9 or later. The news here is its presentation in release 0.40.11, not a claim that d3 launched for the first time anywhere.
Who could use it
Owners of small service businesses with recurring enquiry types and a technical specialist who can configure an integration.
For illustration, a repair shop could define “new job”, “repair status”, “complaint” and “other”. The model only suggests a category; a staff member checks it before routing the enquiry. It does not send a customer reply or modify a job. This is an editorial pilot design, not a measured real-world case.
Availability and limits
Ollama documents support for text, structured state and still images in its d3 package; video is unavailable in this integration. The limitations section specifies a 4,096-token context for these packages: do not assume a long enquiry history will fit without checking. Do not assume that a particular package supports every capability of the original model. The models use the Apache 2.0 licence.
Ollama suggests starting with 4B on a machine with 16 GB of memory. Its published system guidance was tested on Apple Silicon; memory use differs on other systems. This is not our speed test or a compatibility check of your computer. Technical configuration, maintenance and output review still take time.
Before running a pilot
- Choose one workflow and define its categories in writing, including ambiguous cases.
- Prepare de-identified examples in your customers’ languages. Record the correct decisions beforehand.
- Compare suggested categories with a staff member’s decisions. Count misrouted complaints separately.
- Keep a manual path for uncertain cases. Do not turn a high model score into permission to issue refunds or change bookings automatically.
A returned probability does not guarantee the same frequency of correct decisions in your incoming workload. Nadali has not tested d3 and makes no claim about time savings or Ukrainian-language accuracy.
Bottom line: test a narrow routing step, not the replacement of your whole support function. To prepare the pilot, read how to build an AI acceptance test set.
Author: Nadali. Prepared with AI assistance; sources rechecked on 27 October 2026. Cover: an original AI illustration, not a photograph of a real deployment or test.
