vLLM Semantic Router has released Decision 3.0, a new family of five open-source multimodal decision models ranging from 0.8B to 27B parameters. These models can process text, JSON, and images to answer specific questions, returning probabilities rather than generated text, which is useful for routing and guardrail applications. The models are fine-tuned from Qwen base models and offer competitive performance on internal evaluations, with the 9B d3-flash model outperforming the previous generation's 27B model. Latency measurements indicate efficient processing on AMD Instinct MI325X GPUs. AI
IMPACT Provides developers with efficient routing and guardrail capabilities using probabilistic outputs from multimodal inputs.
RANK_REASON New open-source multimodal model family release from a recognized lab. [lever_c_demoted from frontier_release: ic=1 ai=1.0]
- AMD Instinct MI325X
- Apache Software License 2.0
- Decision 3.0
- JPT-0.8B
- Perplexity Decider v1.1
- Qwen
- Torchcast Decision 27B
- transformers
- vLLM Semantic Router
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
