vLLM Semantic Router Releases Decision 3.0 Family of Open Multimodal Decision Models
- Model Size Range: Five distinct sizes—d3-lite (0.85B), d3-nano (2.21B), d3-mini (4.54B), d3-flash (8.39B), and d3 (26.09B)—are built upon Qwen base checkpoints.
- Multimodal Capabilities: Each model features an integrated vision encoder supporting PNG, JPEG, and WebP images up to 1.6 megapixels per image for unified visual-text analysis.
- Probability-Based Output: Instead of text generation, models answer questions as choices, Yes/No, or numeric scores, returning calibrated probabilities for every answer in a single call.
- Benchmark Performance: The flagship 27B model achieved strong internal evaluation results, scoring 97.7 on CORD+FUNSD document extraction and outperforming older generations.
The Context & Background
Built as an open alternative for the 'system one' routing pattern, Decision 3.0 addresses the growing developer need for efficient, programmatic guardrails and classification systems in AI workflows. By providing direct probability scores rather than text output, these models offer faster and more reliable integration for modern applications.
Why it matters:Decision 3.0 transforms how developers handle AI request routing, classification, and guardrails by bypassing traditional text generation in favor of direct probability outputs. This approach drastically reduces parsing overhead, lowers latency, and provides reliable, threshold-ready metrics for enterprise applications.
Original Sources & Coverage (1)
This event has been analyzed and clustered from the following reporting outlets. Click to read their direct coverage:
