d3 is a family of decision models from [vLLM Semantic Router](https://huggingface.co/collections/vllm-sr/decision-30). Give it text, structured state, or a still image, then ask it to choose a catego…
d3 is a family of decision models from [vLLM Semantic Router](https://huggingface.co/collections/vllm-sr/decision-30). Give it text, structured state, or a still image, then ask it to choose a category, answer yes or no, or rate something on a scale. It returns scores and probabilities for the options you provide. It does not write a conversational answer.
Use d3 to route support requests, triage documents, prioritize cases, or check whether an image meets a criterion. You can ask several questions about the same input in one request.
## Get started
`d3` runs the 4B model, a good starting point on a 16 GB machine. This example routes a support request:
```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'
```
d3 requires Ollama 0.40.9 or later.
## Choose a size
| Model | Best fit | Tested Apple Silicon system | Loaded model memory |
| --- | --- | ---: | ---: |
| `d3:2b` | Fast, local routing and classification | 8 GB | 3.62 GB |
| `d3` (4B) | General use | 16 GB | 5.63 GB |
| `d3:9b` | Harder decisions that benefit from more capacity | 32 GB | 10.09 GB |
| `d3:27b` | Highest accuracy in the publisher's family evaluation | 64 GB | 28.61 GB |
Memory figures are the peak loaded model allocations measured in Ollama for the default MXFP8 packages on Apple Silicon with a 4,096-token context. The system recommendations leave room for macOS and the full test workload. On other supported systems, these names select Q8_0 packages and memory use will differ. Format-specific tags are available when you need a particular quantization.
## Quality
The publisher reports these [Jev Decision Index 0.3.2](https://huggingface.co/vllm-sr/d3/blob/cf15549a78fc52598ddd1951f321c58372bbd64f/README.md) scores for the original checkpoints. They are **not** scores for the Ollama quantized packages.
| Model | 2B | 4B | 9B | 27B |
| --- | ---: | ---: | ---: | ---: |
| Text index | 38.9 | 52.6 | 55.8 | 66.3 |
| Vision index | 54.7 | 63.2 | 66.9 | 71.3 |

## Limits
- Ollama supports text, structured state, and still images for d3. Video and Box/Track decisions are unavailable.
- d3 scores the options and levels you supply. Clear, distinct criteria matter, especially when an answer triggers an action.
- Scores and probabilities are model outputs, not guaranteed real-world frequencies. Check performance on examples from your own workflow before setting thresholds.
- These packages use a 4,096-token context.
The models are released under [Apache 2.0](https://huggingface.co/vllm-sr/d3/blob/cf15549a78fc52598ddd1951f321c58372bbd64f/LICENSE). Publisher checkpoints: [2B](https://huggingface.co/vllm-sr/d3-nano), [4B](https://huggingface.co/vllm-sr/d3-mini), [9B](https://huggingface.co/vllm-sr/d3-flash), [27B](https://huggingface.co/vllm-sr/d3).