d3-ultra
A 27B open-weight multimodal decision model released by the vLLM Semantic Router team that adopts a discriminative scoring paradigm, reading text, images and video in a single forward pass and outputting a probability distribution over candidate answers, designed for routing, classification and answer-selection scenarios.
Timeline (last 90 days)
The model weights of d3-ultra are open-sourced on Hugging Face under the Apache-2.0 license.
AMD, in collaboration with the vLLM Semantic Router team, launched the d3-ultra model.
d3-ultra supports per-answer probability output, visual grounding, and video tracking.
d3-ultra is trained and run for inference entirely on AMD Instinct GPUs.
Public quantitative evaluation results are currently lacking on d3-ultra's multimodal dimension.
d3-ultra is a discriminative scoring model, whose typical structure is an encoder or fusion encoder followed by a scoring head that computes a logit for each candidate answer and obtains a probability distribution via softmax.
d3-ultra claims to handle text, images and video uniformly in a single forward pass, which typically means using a unified tokenizer to map different modalities into a shared representation space and having the same Transformer backbone perform cross-attention computation.
Judging from the naming, d3-ultra is a component under the vLLM Semantic Router project rather than an independent foundation model, and it is used by being embedded in the semantic routing pipeline.
The d3-ultra original text does not mention a probability calibration scheme, nor does it explain the actual processing granularity of video input (number of sampled frames, duration limit, whether audio is supported) or the integration approach with the vLLM inference stack.
The 27B parameter count is above average among open-source models, balancing capability and deployment cost.
4 more timeline events
All Facts (14)
d3-ultra supports per-answer probability output, visual grounding, and video tracking.
50%Unverifiedd3-ultra is trained and run for inference entirely on AMD Instinct GPUs.
50%UnverifiedAMD, in collaboration with the vLLM Semantic Router team, launched the d3-ultra model.
50%UnverifiedOfficially, d3-ultra is disclosed to have achieved a score of 87.2 on the team's self-built SR-Text test.
50%UnverifiedPublic quantitative evaluation results are currently lacking on d3-ultra's multimodal dimension.
50%Unverifiedd3-ultra is a discriminative scoring model, whose typical structure is an encoder or fusion encoder followed by a scoring head that computes a logit for each candidate answer and obtains a probability distribution via softmax.
50%Unverifiedd3-ultra claims to handle text, images and video uniformly in a single forward pass, which typically means using a unified tokenizer to map different modalities into a shared representation space and having the same Transformer backbone perform cross-attention computation.
50%UnverifiedThe 27B parameter count is above average among open-source models, balancing capability and deployment cost.
50%UnverifiedJudging from the naming, d3-ultra is a component under the vLLM Semantic Router project rather than an independent foundation model, and it is used by being embedded in the semantic routing pipeline.
50%UnverifiedThe d3-ultra original text does not mention a probability calibration scheme, nor does it explain the actual processing granularity of video input (number of sampled frames, duration limit, whether audio is supported) or the integration approach with the vLLM inference stack.
50%Unverifiedd3-ultra is positioned as a decision model that reads text, images and video in a single pass and returns a probability score for each candidate answer rather than generating natural language text.
50%UnverifiedDetailed information on the new model's specific architecture, training data and benchmark results has not yet been disclosed in the original tweet.
50%UnverifiedThe model weights of d3-ultra are open-sourced on Hugging Face under the Apache-2.0 license.
50%UnverifiedThe vLLM Semantic Router team released a 27B open-weight multimodal decision model named d3-ultra.
50%