It processes text, images, and video, returning classifications, yes/no answers, scores, bounding boxes, and tracking results without generating explanatory text.
In the team's local test, d3-turbo processed 255 questions about one image in 2.12 seconds, compared with 32.94 seconds for d3-flash.
Technical Specifications
| Specification | d3-turbo |
| Parameters | 27B |
| Architecture | Multimodal decision model |
| Input | Text, images, video |
| Output | Choice, yes/no, score, box, track |
| Default context | 32K tokens per question |
| Maximum context | 128K tokens per question |
| Precision | BF16 |
| License | Apache 2.0 |
The 128K context window requires long-context mode.
d3-turbo vs. d3-flash
| Metric | d3-turbo | d3-flash |
| Parameters | 27B | 9B |
| 1 question | 302 ms | 154 ms |
| 16 questions | 426 ms | 2.07 s |
| 64 questions | 867 ms | 8.26 s |
| 255 questions | 2.12 s | 32.94 s |
d3-turbo is 15.5× faster for 255 questions, although d3-flash remains faster for a single question.
The performance improvement comes from the StateOnce architecture, which processes shared input once and reuses it across multiple questions.
Accuracy Compared With d3
| Benchmark | d3-turbo | d3 |
| Public Jev Decision Index | 71.7 | 65.5 |
| Estimated Jev Text Index | 70.2 | 66.3 |
| Estimated Jev Vision Index | 72.9 | 71.3 |
On the public decision suite, d3-turbo improves the original d3 score by 6.2 points.
Multimodal Capabilities
- Classification: Multiple-choice decisions, binary answers, and numerical scoring.
- Image understanding: Questions and answer options can contain images.
- Object localization: Identifies objects using bounding-box coordinates.
- Video tracking: Tracks selected objects across frames.
- Long-context processing: Supports up to 128K tokens per question.
Availability
d3-turbo is available under the Apache 2.0 license, with support for local deployment.
- ModelScope
- Hugging Face
Bottom line: d3-turbo offers substantially faster multi-question inference than d3-flash and higher decision accuracy than the original d3. Its main use cases include large-scale classification, image analysis, and video object tracking.
