vLLM Semantic Router 发布 Decision 3.0 决策模型族
5 个开源多模态决策模型（0.8B–27B），一次调用对 Choice/Yes-No/Score 问题返回概率分数，9B d3-flash 在 Jev Decision Index 上以 59.0 超过 Decision 2.0 的 27B。
vLLM Semantic Router's Decision 3.0, a family of 5 open multimodal decision models from 0.8B to 27B. They read text, JSON and images, then answer typed questions in 1 call, returning a probability for every answer instead of generating text. The core idea is “system one” decisions for routers and guardrails. You pass a state, optional images and named Choice, Yes/No or Score questions. Each question gets its own forward pass, so you get calibrated scores you can threshold, not text you have to parse. The size curve is the real story. The 9B d3-flash scores 59.0 on the Jev Decision Index, beating Decision 2.0’s 27B (55.9). The 27B d3 posts 64.1 on text and 71.6 on vision, ahead of Perplexity Decider v1.1 (62.8 and 70.6). The trade-off: those d3 scores are internal evals, and the 0.8B and 2B models don’t lead every board. Apache-2.0, built on Qwen bases, trained on AMD Instinct MI325X. Median text latency runs from 21.7 ms to 84 ms on 1 GPU. Built for routing, moderation and agent guardrails...... Full analysis: https://www.marktechpost.com/2026/10/11/what-is-decision-3-0-vllm-semantic-routers-new-open-decision-models/ Model: https://huggingface.co/collections/vllm-sr/decision-30 Repo: https://github.com/vllm-project/semantic-router @vllm_project