The vLLM Semantic Router team has announced the launch of Decision 3.0, a family of sophisticated multimodal decision models. These models are engineered to process diverse input types, including textual data, structured JSON, and visual content from images, subsequently providing precise, typed answers with associated probabilities.
Comprising five distinct models, Decision 3.0 ranges in size from a compact 0.8 billion parameters to a robust 27 billion parameters, all released under the permissive Apache-2.0 license. This collection provides developers with an efficient mechanism for swiftly classifying, routing, and gating requests, eliminating the traditional necessity of parsing generated text outputs from large language models.
Deep Dive into Decision 3.0's Architecture and Capabilities
Decision 3.0 models are built upon fine-tuned versions of the Qwen base architecture. The largest model, d3 (27B), leverages Qwen3.8-27B, while its four smaller counterparts utilize Qwen3.5 checkpoints. Every model within the series is equipped with an integrated vision encoder, enabling robust image interpretation. These vision encoders vary in size, from 0.10B in d3-lite to 0.46B in the larger d3-flash and d3 models. The system can process several images simultaneously, accepting PNG, JPEG, or WebP formats, each at resolutions up to 1.6 megapixels, ensuring comprehensive visual context for every query.
The operational design of Decision 3.0 is centered around returning probabilities rather than narrative text. Users provide a 'state' (either text or JSON) and optionally images, alongside specific named questions. These questions can be structured as Choices, Yes/No queries, or Scores on a defined scale. The model efficiently addresses all posed questions in a single inference call, yielding a probability for each potential answer. This 'system one' pattern is particularly advantageous for creating reliable routers and guardrails, offering calibrated scores that can be easily thresholded.
Performance Benchmarks and Speed
According to evaluations presented in the model card, Decision 3.0 demonstrates compelling performance. On internal assessments using the Jev Decision Index 0.3.1 board, the flagship d3 (27B) model achieved a score of 64.1 on text-based tasks and 71.6 on vision tasks, positioning it ahead of competitors like Perplexity Decider v1.1.
A notable achievement for this release is the performance of the 9B d3-flash model, which scored 59.0, surpassing the previous generation's 27B Decision 2.0 model (55.9). Across public suites, the new models exhibit gains ranging from +8.0 to +15.1 over their Decision 2.0 predecessors, depending on the model size.
However, leadership across all benchmarks is not uniform. While d3-mini (4B) and d3-flash (9B) lead both their text and vision categories, d3-lite (0.8B) excels in text but ranks third in vision. Similarly, d3-nano (2B) leads its vision category but trails in text performance against LiquidAI d1-3B.
On specific public vision tasks, d3 achieves high scores of 97.3 on InfographicVQA and 89.9 on Mind2Web, though it shows weaker performance on MMMU-Pro vision and Hateful Memes moderation.
Latency and Efficiency
The models are designed for speed. Median latencies for single requests on an AMD Instinct MI325X GPU demonstrate impressive responsiveness:
- d3-nano: 21.7 ms (text), 92.0 ms (with image)
- d3-lite: 23.5 ms (text), 60.7 ms (with image)
- d3-mini: 28.5 ms (text), 115.0 ms (with image)
- d3-flash: 29.7 ms (text), 171.8 ms (with image)
- d3: 84 ms (text), 342 ms (with image)
These performance metrics highlight Decision 3.0 as a significant advancement for developers seeking efficient and accurate multimodal decision-making capabilities within an open-source framework.
This article is a rewritten summary based on publicly available reporting. For the original story, visit the source.
Source: MarkTechPost
