vLLM Semantic Router releases Decision 3.0 models

vLLM Semantic Router released Decision 3.0, open Apache-2.0 multimodal models that return structured decisions instead of generated text.

vLLM Semantic Router released Decision 3.0, which MarkTechPost describes as a family of five Apache-2.0 multimodal decision models sized from 0.8B to 27B parameters. Hugging Face trending coverage instead focuses on vllm-sr/d3, a 26.09B model with a 0.46B vision encoder that ranked #24 on 10 October 2026. Both sources say the models handle choice, yes/no, and score questions in a single call without generating free-form text; the d3 report specifically says they return probabilities. The sources disagree on modality: d3 is described as accepting text or JSON plus images and videos, whereas MarkTechPost mentions text and images only. On the official Jev Decision Index 0.3.2 text board, d3 scores 66.3, first place and 10.4 above Decision 2.0, with a public-suite score of 65.45. The same report gives an internal Perception Test multiple-choice video QA score of 77.4 and a median text latency of 54.4 ms on one AMD Instinct MI325X.

• vLLM Semantic Router released Decision 3.0 under the Apache-2.0 license.
• MarkTechPost (2026-10-11) describes five multimodal models ranging from 0.8B to 27B parameters.
• Hugging Face coverage (2026-10-10) highlights vllm-sr/d3, a 26.09B model including a 0.46B vision encoder, trending at #24.
• The models answer choice, yes/no, and score questions in one call without generating free-form text; the d3 report says the outputs are probabilities.
• Sources differ on inputs: d3 coverage includes text or JSON plus images and videos, while MarkTechPost mentions text and images only.
• On the Jev Decision Index 0.3.2 text board, d3 scores 66.3, first place and 10.4 above Decision 2.0; the public suite score is 65.45.
• Reported internal Perception Test multiple-choice video QA is 77.4, and median text latency is 54.4 ms on one AMD Instinct MI325X.