AI news ·
Decision 3.0 models skip text generation to deliver direct yes/no and score answers
vLLM Semantic Router released Decision 3.0, five multimodal models from 0.85B to 26.09B parameters under Apache-2.0 that output typed decisions directly instead of parsed prose.
The vLLM Semantic Router team released Decision 3.0, a family of five multimodal decision models spanning 0.85B to 26.09B parameters, all under the Apache-2.0 license. The models accept text, JSON, and images as input and return typed answers-Choice, Yes/No, or Score-as direct probability outputs, eliminating the need to parse generated prose for routing or gating decisions.
For production systems that classify, route, or gate requests, this architecture sidesteps a common bottleneck. Traditional pipelines wait for a language model to generate a sentence, then use brittle regex or parsing logic to extract a verdict. Decision 3.0 models score in a single pass and hand back structured results. The team tested the lineup on one AMD Instinct MI325X GPU in BF16, with no quantized builds listed and no disclosed context length.
What the benchmark scores actually show
Self-reported scores on the internal Jev Decision Index range from 97.7 on a document extraction task down to 34.1 on a harder vision benchmark. The index itself is version 0.3.1, built by the same team releasing the model. The model card also notes that Torchcast Decision 27B outscores d3 on the public suite. That disclosure matters when evaluating claims that Decision 3.0 sits ahead of Perplexity Decider v1.1 and Jev.
The five sizes-from d3-lite at 0.85B up to a 26.09B flagship-give teams room to test whether a smaller variant like the 2.21B d3-nano handles their workload before allocating GPU budget to the largest model. That practical range, paired with the Apache-2.0 license, matters more for production planning than any single leaderboard number.
How the models differ from standard LLM routers
Standard routing setups often involve prompting a chatbot-style LLM, capturing its text response, and post-processing that text to extract a decision. Decision 3.0 skips generation entirely. It reads multimodal input and outputs a probability distribution tied to a typed question. Developers building classification layers or guardrails get a direct signal without the latency and fragility of parsing free-form text.
The absence of a disclosed context length is a gap for anyone routing longer documents or multi-turn JSON payloads. Until that specification is published, teams evaluating the models for document-heavy workflows will need to test limits themselves.
Why this matters for IT and development teams
For engineers building request routers, content classifiers, or safety guardrails, the Apache-2.0 license removes the procurement friction that comes with proprietary model APIs. The size range means you can benchmark a smaller model on your own JSON schemas and image sets before scaling up. The direct probability output cuts out an entire parsing step that introduces latency and error in production pipelines. If you are evaluating classification models for a customer support routing layer or a document triage system, the real test is whether these scores hold up on your data-not on the Jev Decision Index. Professionals looking to build these skills can explore AI Engineering Courses that cover model selection and deployment patterns for production systems.