{"version":1,"type":"story","url":"https://digestai.news/story/vllm-semantic-router-releases-decision-3-0-multimodal-models-with-5-si","json":"https://digestai.news/story/vllm-semantic-router-releases-decision-3-0-multimodal-models-with-5-si.json","markdown":"https://digestai.news/story/vllm-semantic-router-releases-decision-3-0-multimodal-models-with-5-si.md","slug":"vllm-semantic-router-releases-decision-3-0-multimodal-models-with-5-si","headline":"vLLM Semantic Router releases Decision 3.0 multimodal models with 5 sizes under Apache-2.0","summary":"vLLM Semantic Router has launched Decision 3.0, a family of open-source multimodal decision models that process text, JSON, and images to answer structured questions. The suite includes five model sizes—d3-lite (0.85B), d3-nano (2.21B), d3-mini (4.54B), d3-flash (8.39B), and d3 (26.09B)—all licensed under Apache-2.0. Unlike traditional LLMs, these models return probabilities for predefined choices (Yes/No, Score, or Choice) rather than text, enabling faster classification, routing, and request gating without parsing outputs.\n\nThe models are fine-tuned on Qwen checkpoints (Qwen3.5 for smaller sizes, Qwen3.8-27B for the largest) and include vision encoders for image input. Performance varies by task: the 27B model leads internal benchmarks (64.1 on text, 71.6 on vision), while the 9B d3-flash outperforms Decision 2.0’s 27B in both domains. Latency ranges from **21.7 ms** (text-only) to **342 ms** (with images) on a single AMD Instinct MI325X GPU. Scores on public benchmarks are mixed, with strengths in document extraction (97.7 on KIE) but weaker performance on moderation tasks (44.6 on Hateful Memes). All models are available on Hugging Face and GitHub, with installation notes specifying `transformers==5.17.0` and optional optimizations.","keyPoints":["Five open-source Decision 3.0 models (0.85B–26.09B) support text, JSON, and images under Apache-2.0","d3-flash (9B) outperforms Decision 2.0’s 27B model in text/vision benchmarks per internal evals","Latency ranges from **21.7 ms** (text) to **342 ms** (with images) on AMD Instinct MI325X"],"whyItMatters":"Decision 3.0’s structured outputs and open licensing could accelerate routing systems for apps needing fast, probabilistic decisions—e.g., moderation, document classification, or API gateways—without relying on proprietary models.","category":{"slug":"models","name":"Generative AI & Models","url":"https://digestai.news/category/models"},"entities":{"companies":["vLLM Semantic Router"],"models":["Decision 3.0","d3-lite","d3-nano","d3-mini","d3-flash","d3"],"people":[]},"firstPublishedAt":"2026-10-11T21:43:41Z","updatedAt":"2026-10-11T21:43:41Z","sourceCount":1,"hasPrimarySource":false,"sources":[{"outlet":"MarkTechPost","title":"What is Decision 3.0? vLLM Semantic Router’s New Open Decision Models","url":"https://marktechpost.com/2026/10/11/what-is-decision-3-0-vllm-semantic-routers-new-open-decision-models","publishedAt":"2026-10-11T21:43:41Z","type":"press","primary":false,"lead":true}],"sourceNotes":null,"discussions":[],"thread":null,"cite":{"text":"Digest AI, \"vLLM Semantic Router releases Decision 3.0 multimodal models with 5 sizes under Apache-2.0\", 11 October 2026, https://digestai.news/story/vllm-semantic-router-releases-decision-3-0-multimodal-models-with-5-si","publisher":"Digest AI","title":"vLLM Semantic Router releases Decision 3.0 multimodal models with 5 sizes under Apache-2.0","datePublished":"2026-10-11T21:43:41Z","url":"https://digestai.news/story/vllm-semantic-router-releases-decision-3-0-multimodal-models-with-5-si"},"generatedBy":"Written by Digest AI's editorial model from the linked sources; the sources are the record.","license":"Headlines, digests and key points are written by Digest AI and may be quoted with a link to the story page. Linked articles belong to their publishers. Terms: https://digestai.news/terms#reuse"}