d1-3B reframes some common production needs: instead of generating free-form text, it emits a calibrated decision distribution about a presented state (text, JSON, images, or a mix) in one forward pass. That single-pass, probability-first design makes it suitable where a deterministic, auditable decision or routing choice is the desired output rather than narrative or explanation.
Key Capabilities
- Structured decisions, not text: answers three question types—noul (binary probability), choice (named-option distribution + confidence), and score (ordered rubric with expected level and confidence). This makes integration into rule systems, routing, moderation, and agent guardrails straightforward.
- Multimodal, long-context backbone: ~3.12B parameters with a vision encoder and 32,768-token context window, so it can reason over lengthy states and images together for a single decision.
- Calibrated probabilities and speed: returns probabilities (no token generation) and is measured for edge and GPU: e.g., single-question latencies under 10 ms on high-end GPUs and under 50 ms on many embedded devices, enabling real-time pipelines.
- Designed for decision pipelines: common uses include intent/topic classification, triage/routing, moderation gates, reranking, LLM-as-judge scoring, and visual inspections; it intentionally avoids free-form generation.
Who it's for and tradeoffs
Great fit if you need deterministic, auditable decisions from multimodal inputs with tight latency or edge-resource constraints — e.g., automated triage, moderation filters, model routing, or tool-call approval in agents. The probability outputs simplify downstream thresholds and monitoring.
Look elsewhere if your application requires open-ended generation, detailed explanations, or conversational AI: this model does not produce text responses and is not optimized for creative or explanatory outputs. Also, while Decision Index and benchmark numbers are strong for its size, accuracy varies by task and you should validate calibration and label definitions for your domain before relying on automated gating.
