What is Decision 3.0? vLLM Semantic Router’s New Open Decision Models
Key Points
- 1The vLLM Semantic Router team has released Decision 3.0, a collection of five open-source multimodal models ranging from 0.8B to 27B parameters designed for efficient classification and routing.
- 2These models improve upon previous generations by returning calibrated probabilities for specific tasks like JSON extraction and image analysis, effectively acting as "system one" routers without needing to parse generated text.
- 3Benchmarked on AMD Instinct MI325X GPUs, the lineup offers significant performance gains, with the 9B model outperforming the 27B version of its predecessor, Decision 2.0.
The Decision 3.0 family, released by the vLLM Semantic Router team, represents a series of five open-source multimodal decision models (ranging from 0.8B to 27B parameters) designed to act as high-efficiency routers, classifiers, and guardrails. Unlike generative large language models (LLMs) that produce text, these models are engineered to return calibrated probabilities for predefined questions, effectively acting as a "system one" pattern for AI pipelines.
Core Methodology and Architecture
The models function by accepting a state (provided as text or JSON) alongside optional multimodal inputs (PNG, JPEG, or WebP images). Each question posed to the model is categorized as a choice, a binary (Yes/No), or a numerical score. Crucially, the model performs a forward pass for each specific question against the input data, outputting a probability distribution that can be thresholded by developers to automate decision-making without the overhead of parsing natural language generation.The architectural foundation relies on Qwen base models:
- d3 (27B): Fine-tuned from Qwen3.8-27B.
- d3-lite, d3-nano, d3-mini, and d3-flash: Based on Qwen3.5 checkpoints (0.8B to 9B).
Integration of vision capabilities is achieved through dedicated vision encoders, which scale with the model size: 0.10B for the smallest variant up to 0.46B for the 8.39B and 26.09B models. These encoders allow for the analysis of images at up to 1.6 megapixels. The technical stack requires , with optional flash-linear-attention support to optimize linear-attention layers.
Performance and Benchmarking
Decision 3.0 is evaluated using the "Jev Decision Index 0.3.1," a framework for ranking decision-making systems. According to internal evaluations, the 27B model outperforms competitors like Perplexity Decider v1.1 and the older Decision 2.0. Notably, the 9B model (d3-flash) demonstrates significant efficiency gains, outperforming the previous generation's 27B model with a score of 59.0 versus 55.9.Performance metrics across tasks are as follows:
- Document Extraction (KIE): The 27B model achieves a peak score of 97.7.
- Vision Tasks: The 27B model reaches 71.6 on the vision board, with specific successes in InfographicVQA (97.3) and Mind2Web (89.9).
- Latency: On a single AMD Instinct MI325X GPU, text-only latency ranges from 21.7 ms (d3-nano) to 84 ms (d3). Multimodal latency with image input ranges from 60.7 ms to 342 ms, respectively.
Technical Specifications Summary
| Model | Parameters | Base Model | Vision Encoder |
| :--- | :--- | :--- | :--- |
| d3-lite | 0.85B | Qwen3.5 | 0.10B |
| d3-nano | 2.21B | Qwen3.5 | 0.33B |
| d3-mini | 4.54B | Qwen3.5 | 0.33B |
| d3-flash | 8.39B | Qwen3.5 | 0.46B |
| d3 | 26.09B | Qwen3.8-27B | 0.46B |
All models are released under the Apache-2.0 license, providing a scalable, low-latency alternative for production environments that require reliable classification rather than generative output.