Why this matters
Most multimodal models focus on generating text; d3 is designed to make structured decisions instead. It accepts text, JSON, multiple images and videos in one request, answers many question types (choice, yes/no, score) in a single forward pass per question, and returns calibrated probabilities for every option — a workflow that fits routing, triage and automated decision applications more directly than free-form text generation.
Key Capabilities
- Decision-first API: returns a probability distribution for each option rather than a free-text reply, simplifying downstream decision-making and metric collection.
- Multimodal inputs: accepts several images (up to 1.6 MP each) and videos (read at 2 fps, up to 32 frames) alongside text or JSON; every question sees all provided media.
- Size and architecture: ~26.09B parameters including a 0.46B vision encoder; built on Qwen/Qwen3.8-27B checkpoints with a structured decision readout head (safetensors). This yields compact, deterministic decision outputs rather than generated text tokens.
- Benchmarks and evaluation: Jev Decision Index 0.3.2 = 66.3 (first on the official text board, public suite 65.45) and Perception Test (video QA) 77.4 on 19,140 questions; shows strong performance on both text and vision decision tasks.
- Production-ready serving: integrates with vLLM and vLLM Semantic Router. Reported throughput examples: 27 requests/s with p99 < 1s on an AMD Instinct MI325X (text, one-question workloads); typical latencies: ~54 ms p50 for single-question text, ~290 ms p50 for one image.
- Structured-question types: supports up to 255 choice options, yes/no (noul), and score questions; returns per-option probabilities and a confidence value for the chosen option.
Who it's for and trade-offs
Great fit if you need a model that directly answers structured decision queries about mixed media inputs and you plan to deploy at scale using GPU-backed serving (vLLM / vLLM-sr). It is especially useful for automated routing, moderation triage, form/document verification with receipts/images, and multiple-choice video QA.
Look elsewhere if you need free-form text generation, extremely small-edge deployment (d3 requires substantial GPU memory), or guaranteed deterministic byte-for-byte answers across different runtimes without careful kernel and batching configuration. d3 is licensed Apache-2.0 and is built on Qwen weights, so follow downstream licensing and deployment constraints for production use.
Where it fits
Positioned as a decision-centric multimodal foundation model rather than a general-purpose chat LLM. Compared with its Decision 2.0 predecessor, it reports significant gains on the Jev Decision Index and public suites, and among Decision 3.0 variants it sits at the top of the official text and vision Pareto frontiers.
