vLLM releases d3-turbo 27B, a router model that returns probabilities, not text

The vLLM Semantic Router team published d3-turbo on Hugging Face: 26.09B parameters plus a 460M vision encoder under Apache 2.0. Instead of generating sentences, the model returns probabilities for predefined options — choice, yes/no, score, box and track. It is 15.5x faster than d3-flash at 255 questions per image, but twice as slow on a single question.

• 26.09B parameters plus a 460M vision encoder, Apache 2.0 license
• At 255 questions d3-turbo answers in 2.12 s versus 32.94 s for d3-flash
• On a single question turbo is slower: 302 ms versus 154 ms
• Context up to 131,072 tokens with long_context=True, 32,768 by default