# d3-flash

> d3-flash is a 9B multimodal decision model from vLLM Semantic Router (Decision 3.0). Send a state, optionally with images and videos, plus typed questions (choice, noul, score and box) and get a calibrated probability for each answer in one forward pass, with no generated text. Official Jev Decision Index scores are 54.27 on the text board and 66.89 on the vision board. Box answers come from grid scoring over the image, which costs roughly 18,000 input tokens per box question. Context is 131,072 tokens per request. Call POST /api/v1/decisions (POST /api/v1/systemone is the same handler). The service logs requests and may train on them, so it is skipped when Always enforce ZDR is on (Dashboard, Privacy).

**ID**: `vllm-sr/d3-flash`  
**Creator**: vllm-sr  
**Category**: text  
**Context**: 131K tokens  
**Input**: $0  
**Output**: $0  
**Released**: 2026-10-10  
**Web page**: https://anyrouter.dev/model/vllm-sr/d3-flash

**Input modalities**: text, image, video  
**Output modalities**: text  

**Capabilities**: systemone

**Supported parameters**: state, model, questions

## Usage

**Endpoint**: `POST https://anyrouter.dev/api/v1/decisions`

```bash
curl https://anyrouter.dev/api/v1/decisions \
  -H "Authorization: Bearer $ANYROUTER_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
  "state": "Help! My payouts have been failing for 3 days.",
  "model": "vllm-sr/d3-flash",
  "questions": {
    "is_urgent": {
      "type": "noul",
      "instructions": "Does this convey urgency?"
    }
  }
}'
```

Decisions is not chat: send `state` and typed `questions` (noul / choice / score). Requires a BYOK key for TypeSafe, LLM Gateway, Vercel AI Gateway, AIHubMix, AutoJev, or Fastino.

## Providers

| Provider | Input | Output |
| --- | --- | --- |
| vllm-sr | $0 | $0 |
| vllm-sr-byok | $0 | $0 |

## Model card

- Total parameters: 8.39B [1]
- Native context: 131072 tokens [2]
- License: Apache-2.0, open weights (https://huggingface.co/vllm-sr/d3-flash) [1]

### Strengths [1,2]

- Official Jev Decision Index scores of 54.27 (text board) and 66.89 (vision board)
- Answers every typed question about one input in a single forward pass, with a probability per answer
- Reads text, images and videos

### Known limits [2]

- Generates no text, so there is no written rationale
- A request may take at most 30 seconds, with per-key rate limits and daily quotas
- Requests and responses may be logged and used to train models

### Use cases [1]

- Routing, classification and moderation of text, images and video
- Scoring and grading with calibrated probabilities

### API notes [2]

- System One request shape: a state plus named questions of type choice, noul, score and box
- Answers come back keyed by question name, with input token usage and zero output tokens

### Links

- model card: https://huggingface.co/vllm-sr/d3-flash
- docs: https://decision.vllm-sr.ai/docs

### Sources

1. [d3-flash model card](https://huggingface.co/vllm-sr/d3-flash) (accessed 2026-10-28)
2. [Decision API reference](https://decision.vllm-sr.ai/docs) (accessed 2026-10-28)

## Source

- Upstream docs: https://decision.vllm-sr.ai/docs