Decision 3.0
An open-source multimodal decision model series released by vLLM Semantic Router under the Apache-2.0 license. It includes five model sizes ranging from 0.8B to 27B parameters and focuses on structured decision outputs—multiple choice, yes/no, and scoring tasks—rather than generating free-form natural language text.
Timeline (last 90 days)
Decision 3.0 models are available on Hugging Face (vllm-sr/decision-30 collection).
Decision 3.0 took first place on both the text and vision leaderboards of the Jev Decision Index.
Multiple questions for the same input can be completed in a single call, with each question performing one forward pass on the input.
The parameter counts published for Decision 3.0 include the vision encoder, with the vision encoder accounting for 0.33B in the 4B version and 0.46B in the 27B version.
Decision 3.0 image input supports multiple images per request, with a maximum resolution of 1.6 MP.
Decision 3.0 model weights use BF16 precision, can be loaded via the Transformers library (requiring trust_remote_code=True), and training was completed on AMD Instinct MI325X accelerator cards.
The 27B version of Decision 3.0 scored 64.7 on the text leaderboard and 71.6 on the vision leaderboard, leading all participants on both leaderboards.
The 9B version of Decision 3.0 (codenamed d3-flash) outperformed the previous generation's 27B model, Decision 2.0, in evaluations.
No Decision 3.0 model was exposed to video data during training, yet on all 19,140 questions in the Perception Test validation set, the 27B version achieved a zero-shot score of 77.4 and the 4B version scored 70.3.
The vLLM Semantic Router team released a family of decision models called Decision 3.0 (abbreviated as d3), covering multiple parameter scales from 0.6B to 27B, released under the Apache-2.0 open-source license, and supporting three input modalities: text, image, and video.
8 more timeline events
All Facts (17)
The 9B version of Decision 3.0 (codenamed d3-flash) outperformed the previous generation's 27B model, Decision 2.0, in evaluations.
50%UnverifiedNo Decision 3.0 model was exposed to video data during training, yet on all 19,140 questions in the Perception Test validation set, the 27B version achieved a zero-shot score of 77.4 and the 4B version scored 70.3.
50%UnverifiedDecision 3.0 took first place on both the text and vision leaderboards of the Jev Decision Index.
50%UnverifiedMultiple questions for the same input can be completed in a single call, with each question performing one forward pass on the input.
50%UnverifiedThe parameter counts published for Decision 3.0 include the vision encoder, with the vision encoder accounting for 0.33B in the 4B version and 0.46B in the 27B version.
50%UnverifiedDecision 3.0 image input supports multiple images per request, with a maximum resolution of 1.6 MP.
50%UnverifiedDecision 3.0 model weights use BF16 precision, can be loaded via the Transformers library (requiring trust_remote_code=True), and training was completed on AMD Instinct MI325X accelerator cards.
50%UnverifiedThe 27B version of Decision 3.0 scored 64.7 on the text leaderboard and 71.6 on the vision leaderboard, leading all participants on both leaderboards.
50%UnverifiedDecision 3.0 can answer three types of structured questions: multiple choice (choice), yes/no, and scoring (score)
50%UnverifiedThe core positioning of Decision 3.0 is not text generation but rather a focus on making decisions without generating text
50%UnverifiedDecision models reframe tasks as classification or scoring problems, directly producing structured results without generating descriptive text
50%UnverifiedSmaller models (such as 0.8B) are suitable for edge devices or latency-sensitive real-time routing scenarios, while larger models (27B) are suitable for tasks requiring higher judgment accuracy
50%UnverifiedThe Apache-2.0 license allows enterprises and developers to relatively freely integrate, modify, and distribute these models in commercial products
50%UnverifiedDecision 3.0 can process text and images simultaneously in a single call
50%UnverifiedDecision 3.0 models are available on Hugging Face (vllm-sr/decision-30 collection).
50%UnverifiedThe vLLM Semantic Router team released a family of decision models called Decision 3.0 (abbreviated as d3), covering multiple parameter scales from 0.6B to 27B, released under the Apache-2.0 open-source license, and supporting three input modalities: text, image, and video.
50%UnverifiedvLLM Semantic Router released Decision 3.0, a set of multimodal decision models licensed under Apache-2.0
50%