# d3 on Ollama: the decision model family with vision at every size

vLLM Semantic Router's d3 landed on the Ollama library today: four decision models that score your options instead of chatting — and, as far as we've seen, the first of this autumn's decision-model wave with vision at every size.

Published 2026-10-30 · https://inpherio.co.uk/blog/d3-on-ollama-the-decision-model-family-with-vision-at-every-size

Key points:
- d3 is a family of decision models from vLLM Semantic Router, on Ollama from 30 October 2026 under Apache 2.0 — it scores choices, yes/no answers and ratings instead of writing prose.
- Four sizes (2b, 4b, 9b, 27b), all with vision and a 256K window; Ollama's measured loads run 3.62GB to 28.61GB on Apple Silicon, from an 8GB Mac up to a 64GB Mac.
- Run it with `ollama run d3` (the 4B, good on a 16GB machine); you need Ollama 0.40.9 or later.
- Publisher-reported Jev Decision Index scores climb from 38.9 (2B text) to 66.3 (27B text) — original checkpoints, not the Ollama quants, and no independent benchmarks yet.
- Among the decision models we've covered this autumn, d3 stands out for image input at every size and a 256K context window; limits are no video, 4,096-token packages, and probabilities you must calibrate yourself.

## What d3 is

If you've been following this space since October, you'll have noticed a new category appearing on the Ollama library: decision models that don't chat at all. [d3](https://ollama.com/library/d3), which landed today (30 October 2026), is the latest — and, as far as we've seen, the first of this autumn's decision models with vision built in across every size.

d3 is a family of decision models from vLLM Semantic Router, released under Apache 2.0. The idea is simple: give it text, structured state, or a still image, then ask it to pick a category, answer yes or no, or rate something on a scale. It returns scores and probabilities for the options you supply. It never writes a conversational answer.

The use cases the page lists are exactly what you'd expect: routing support requests, triaging documents, prioritising cases, or checking whether an image meets a criterion. You can ask several questions about the same input in a single request.

It joins a busy autumn lineup — [kev](/blog/kev-on-ollama-jared-palmer-s-decision-model-family-hands-on), [pplx-decider](/blog/pplx-decider-on-ollama-perplexity-s-27b-model-that-never-writes-a-word), [Clef Flash](/blog/clef-flash-on-ollama-cloudflare-s-9b-decision-model-hands-on), [Strands Decider](/blog/strands-decider-on-ollama-a-2b-model-that-answers-with-probabilities-n) and [nimble-v3](/blog/nimble-v3-on-ollama-a-9b-decision-model-that-answers-with-probabilitie), all of which we've covered over the past few weeks. d3's differentiator among that group: every size — 2B through 27B — accepts images, not just text.

## How to run it

The default tag runs the 4B model, which Ollama calls "a good starting point on a 16 GB machine":

```sh
ollama run d3
```

You'll need Ollama 0.40.9 or later.

The readme's routing example gives you the flavour of the interface:

```sh
ollama run d3 "A customer says their card was charged twice." \
  --ask "Which team should handle this?" \
  --choice 'billing=Invoices and charges' \
  --choice 'returns=Refunds and replacements' \
  --choice 'technical=App errors'
```

One input, one question, three labelled choices — and you get scores and probabilities back rather than prose.

As of release day, the Ollama library page is the only documentation we've seen for d3. There are no LM Studio, llama.cpp or vLLM instructions on it. If you're weighing up engines generally, our [LM Studio vs Ollama comparison](/blog/lm-studio-vs-ollama-which-one-should-you-actually-install) covers the trade-offs; for now, Ollama is the only documented way to run d3.

## Sizes and memory: what each one needs

There are four sizes: 2b, 4b (the default), 9b and 27b. Downloads run from 3.1GB for the 2b up to 29–30GB for the 27b. All four accept text and images, and all list a 256K context window.

Ollama has published measured loaded-model memory on Apple Silicon, using the default MXFP8 packages at a 4,096-token context:

| Tag | Download | Loaded memory (Apple Silicon) | Tested system | Best fit |
|---|---:|---:|---:|---|
| d3:2b | 3.1–3.2GB | 3.62GB | 8GB Mac | Fast, local routing and classification |
| d3 (4B) | 5.6–5.8GB | 5.63GB | 16GB Mac | General use |
| d3:9b | 10GB | 10.09GB | 32GB Mac | Harder decisions needing more capacity |
| d3:27b | 29–30GB | 28.61GB | 64GB Mac | Highest accuracy in the publisher's evaluation |

The system recommendations deliberately leave room for macOS and the full test workload.

A wrinkle for non-Apple hardware: Ollama only publishes Apple Silicon figures, and on other supported systems these tag names select **Q8_0** packages — so expect a larger download and more VRAM than the MXFP8 figures above suggest. Format-specific tags exist if you need a particular quant. As a rough, unverified rule, budget the download size plus overhead; a 12GB gaming GPU is a plausible floor for the 4B, but that's our extrapolation, not a measured figure.

Translated into hardware classes, with that caveat in mind:

- **2b** fits most laptops, including 8GB machines.
- **4b** suits a 16GB Mac; on a GPU, a 12GB card is the plausible floor.
- **9b** wants a 32GB Mac, or a GPU with room for a ~10GB model plus context.
- **27b** is server or 64GB Mac territory.

One more caveat worth flagging: although the models list a 256K context window, these packages ship configured with a 4,096-token context. That's fine for routing and triage, but the Ollama page doesn't say how — or whether — you can raise the package context, so expect to configure it yourself if you need more. And if you do push context up, budget more memory: KV cache grows with context, and 256K on a 27B model is not a small amount.

## How good is it?

The publisher reports Jev Decision Index 0.3.2 scores for the original checkpoints:

| | 2B | 4B | 9B | 27B |
|---|---:|---:|---:|---:|
| Text index | 38.9 | 52.6 | 55.8 | 66.3 |
| Vision index | 54.7 | 63.2 | 66.9 | 71.3 |

Two caveats the page is upfront about, and we'll repeat: these are **not** scores for the Ollama quantised packages, and they're the publisher's own numbers. We've seen no independent benchmarks — the same situation as kev, pplx-decider and bonsai-2 at release.

The pattern across sizes is interesting. The biggest step on text is 2B→4B (+13.7 points, 38.9 → 52.6); the 9B→27B step is the next one worth paying for (+10.5, 55.8 → 66.3). The 4B→9B step is marginal (+3.2) — if you're between sizes, that's the upgrade to skip. Don't read too much into vision scoring higher than text at every size, though — the page doesn't say the two indices are on comparable scales, so we won't claim vision is "better" than text.

## d3 vs the other decision models we've covered

Here's how d3 sits against the decision models we've written about this autumn. One note before the table: this reflects our launch coverage of each model, not a fresh benchmark — follow the links for the details we reported at the time.

| Model | Maker | Sizes | Vision | What we noted at launch |
|---|---|---|---|---|
| d3 | vLLM Semantic Router | 2B–27B | Yes, all sizes | 256K context; images everywhere |
| [kev](/blog/kev-on-ollama-jared-palmer-s-decision-model-family-hands-on) | Jared Palmer | 0.8B–27B | No | Text-only; Apple Silicon MLX runner, so Mac-only |
| [pplx-decider](/blog/pplx-decider-on-ollama-perplexity-s-27b-model-that-never-writes-a-word) | Perplexity | 27B, plus a 4B flash sibling | Flash only | 17GB download for the 27B |
| [Clef Flash](/blog/clef-flash-on-ollama-cloudflare-s-9b-decision-model-hands-on) | Cloudflare | 9B | No | Single mid-size option |
| [Strands Decider](/blog/strands-decider-on-ollama-a-2b-model-that-answers-with-probabilities-n) | Strands Agents | 2B | No | 4,096-token context window |
| [nimble-v3](/blog/nimble-v3-on-ollama-a-9b-decision-model-that-answers-with-probabilitie) | Bespoke Labs | 9B | No | Single mid-size option |

Of that group, d3 is the only one with image input at every size, and the only one with a 256K context window. Strands Decider's 4,096-token limit confines it to short documents; d3's packages also ship at 4,096, but the underlying models list a 256K window — whether you can actually use it in Ollama is something to test before you rely on it.

kev remains the pick if you're on a Mac and want the smallest footprint — its 0.8B checkpoint is smaller than anything d3 offers, though it's text-only and Mac-only. pplx-decider is the heaviest single option we've covered at 17GB; d3:27b at 29–30GB now takes the heavyweight crown, but it offers a 2B escape hatch at the other end of the family.

The honest summary: none of these have independent benchmarks yet. Treat every number in this space as vendor-reported until that changes.

## Limits to know before you wire it into anything

The page is unusually clear about limits, which we appreciate:

- **No video.** Ollama supports text, structured state and still images; Box/Track decisions are unavailable.
- **It scores what you supply.** d3 evaluates the options and levels you give it. Clear, distinct criteria matter — especially when an answer triggers an action downstream.
- **Probabilities aren't guarantees.** Scores and probabilities are model outputs, not calibrated real-world frequencies. Check performance on examples from your own workflow before setting thresholds.
- **4,096-token packages.** Fine for routing and triage; not for long documents out of the box.
- **It never writes prose.** If your workflow needs an explanation alongside the decision, you still want a chat model — d3 makes the call, something else explains it. A small local model like [llama3.2](/models/llama3.2:latest) covers that side cheaply.

(Live on Inpherio: https://inpherio.co.uk/models/llama3.2:latest)

## Worth a look if you're routing or triaging locally

If you have a 16GB Mac, `ollama run d3` is a five-minute experiment, and the 4B is the sensible default. The 2B at 3.62GB loaded is small enough to sit permanently resident as a router in front of your other local models — that's arguably its most compelling role in a homelab.

What we'd pick: the 4B for general use, the 2B as an always-on router, and the 27B only if you're already running hardware that can hold it and the 9B-to-27B quality step matters for your workload.

If you'd rather try the decision-model pattern through an API before committing hardware, or you have a machine that sits idle between experiments, the marketplace covers both:



---

*For more on this space, see everything we've written in [new models](/blog/tag/new-models) and [local AI](/blog/tag/local-ai).*

## FAQ

**What is d3 on Ollama?** d3 is a family of decision models from vLLM Semantic Router, released on the Ollama library on 30 October 2026 under Apache 2.0. It doesn't chat — it returns scores and probabilities for options you supply, given text, structured state or a still image.
**How much memory does d3 need?** Ollama's measured loads on Apple Silicon (MXFP8, 4,096-token context) are 3.62GB for 2b, 5.63GB for 4b, 10.09GB for 9b and 28.61GB for 27b. Tested systems run from an 8GB Mac to a 64GB Mac.
**What's the difference between d3:2b, d3:4b, d3:9b and d3:27b?** All four accept text and images with a 256K window; they differ in capacity and memory. Publisher-reported text scores climb 38.9 → 52.6 → 55.8 → 66.3, with the largest step from 2B to 4B (+13.7).
**Does d3 support images?** Yes — every size accepts still images alongside text and structured state. Video and Box/Track decisions are unavailable in Ollama.
**Can d3 write normal answers?** No. It returns scores and probabilities for the options you provide and never writes prose. If your workflow needs an explanation alongside the decision, you still want a chat model.
**Which Ollama version do I need for d3?** Ollama 0.40.9 or later.
