Handle Text, Image, and Video Decisions in One Go: vLLM Releases d3-ultra 27B Open-Weight Multimodal Routing Model
The vLLM Semantic Router team has launched the d3-ultra multimodal decision model with 27 billion parameters, which natively supports three input types: text, images, and video. It outputs the corresponding probabilities for all candidate results in a single inference pass and scores 87.2 on the team's self-developed SR-Text semantic routing benchmark. The model is open-weight licensed, supports local private deployment, simplifies the decision-making pipeline for multimodal AI applications, and reduces latency and computing costs.
On October 26, 2026, the vLLM Semantic Router team released the d3-ultra model, an open-weight multimodal decision model with 27 billion parameters, as first reported by MarkTechPost, with author Asif Razzaq.
For developers who have built large language model (LLM) applications and Agent systems, decision routing is a core component at the request entry layer. It is responsible for determining the type and complexity of user requests and matching them to corresponding downstream processing resources. Simple small-talk requests are routed to low-cost small models; complex inference requests are sent to flagship large models; requests requiring information retrieval or tool calling are directed to the corresponding functional modules. Smart customer service and AI assistants commonly used by end users mostly have such a routing layer behind them, which directly determines the response speed and overall usage cost of requests.
Most mainstream routing solutions previously only supported plain text input. For multimodal requests involving images and videos, developers needed to first call vision understanding models to convert media content into text descriptions, then send these descriptions along with user questions to text routing for judgment. This entire pipeline requires at least two model calls, leading to higher latency. The process of converting vision to text may also lose key information, affecting the accuracy of final decisions.
d3-ultra integrates multimodal understanding and decision-making capabilities into one model. It can accept three types of input—text, images, and video—in one pass. After inference, it directly provides the corresponding probabilities for each candidate result. Developers can set distribution thresholds directly based on probability values, without needing to build additional preprocessing pipelines for media content.
According to test results released by the team, d3-ultra scored 87.2 on the self-developed SR-Text semantic routing benchmark. This test set covers common routing scenarios such as daily request classification, tool calling judgment, and complexity grading, with the score corresponding to the model's decision accuracy.
As an open-weight model, developers can freely download the complete model weight files and support local private deployment. For enterprise scenarios with data compliance requirements, or online AI products pursuing low latency, deploying the routing model locally avoids transmitting user request content externally, and also eliminates the network overhead of public network API calls.
Source: vLLM Semantic Router Releases d3-ultra: A 27B Open Multimodal Decision Model
发布时间: 2026-10-27 13:50