vLLM Semantic Router releases Decision 3.0, five open multimodal decision models

Decision 3.0 includes d3-lite at 0.85B parameters, d3-nano at 2.21B, d3-mini at 4.54B, d3-flash at 8.39B and d3 at 26.09B. The context length was not disclosed. The models use BF16 weights, no official quantized variants are listed, and latency was measured on one AMD Instinct MI325X GPU. The team said all models were trained on AMD Instinct MI325X GPUs.

Each model is a fine-tune of a Qwen base. The 27B d3 is built on Qwen3.8-27B, while the four smaller models are built on Qwen3.5 checkpoints at 0.8B, 2B, 4B and 9B. Every size carries a vision encoder: 0.10B in d3-lite, 0.33B in d3-nano and d3-mini, and 0.46B in d3-flash and d3. Requests accept several PNG, JPEG or WebP images, each read at up to 1.6 megapixels. Every question sees all images, and each question gets its own forward pass over the input. Installation targets transformers==5.17.0, an optional flash-linear-attention package speeds up the linear-attention layers, and loading requires trust_remote_code=True.

The model card cites the Jev Decision Index 0.3.1, a board that ranks open reproductions of TypeSafe's Jev decision system. The d3 numbers are the team's internal evaluation. On text, d3 scores 64.1, ahead of Perplexity Decider v1.1 at 62.8 and Jev at 60.1. One caveat is that Torchcast Decision 27B posts a higher public-suite score, 65.1 versus 64.9. On the vision board, d3 scores 71.6 against 70.6 for Perplexity Decider v1.1. The highest reported score is 97.7 on KIE (CORD+FUNSD) document extraction for d3, while the lowest is 34.1 on R-Bench-M for d3.

The size story is the strongest part, according to the report. d3-flash, the 9B model, scores 59.0, above Decision 2.0's 27B at 55.9. On the public suite, gains over Decision 2.0 range from +8.0 for the 27B model to +15.1 for the 0.8B model. Class leadership is not uniform: d3-mini and d3-flash lead both their text and vision tables. d3-lite leads on text but ranks third on vision at 41.1, behind JPT-0.8B at 42.1. d3-nano leads its vision table but trails LiquidAI d1-3B on text, 35.8 versus 40.1. On public vision tasks, d3 scores 97.3 on InfographicVQA and 89.9 on Mind2Web, both approximate rebuilds. It is weaker on MMMU-Pro vision at 46.2 and Hateful Memes moderation at 44.6.

The reported median latencies for one request at a time on one AMD Instinct MI325X are 21.7 ms for d3-nano on text and 92.0 ms with an image; 23.5 ms and 60.7 ms for d3-lite; 28.5 ms and 115.0 ms for d3-mini; 29.7 ms and 171.8 ms for d3-flash; and 84 ms and 342 ms for d3.

In the comparison provided, Featured3 from vLLM SR has 26.09B parameters, Perplexity Decider v1.1 has 26B in its Hugging Face listing, Decision 2.0 Vega-27B has 29.37B and LiquidAI d1-3B has 3.12B. Their base models are listed as Qwen3.8-27B, Qwen3.8-27B, Qwen3.8-27B and LFM2.5-VL-3B. Inputs are text, JSON and images for Featured3; text and images for Perplexity Decider v1.1; text and JSON for Decision 2.0 Vega-27B; and text, JSON and images for LiquidAI d1-3B. Context is not disclosed for Featured3, 8,192 tokens per decision for Perplexity Decider v1.1 and 32,768 tokens for the other two. Licenses are Apache-2.0, Apache-2.0, Apache-2.0 and LFM 1.0. The Decision Index on their own cards is 64.1 for Featured3 on v0.3.1 internal, 61.56 for Perplexity Decider v1.1 with version not stated, 56.5 for Decision 2.0 Vega-27B on v0.2.1 and 48.57 for LiquidAI d1-3B on v0.2.1. On the Index 0.3.1 board, the figures are 64.1, 62.8, 55.9 and 40.1. On the vision board 0.3.1, Featured3 has 71.6 and Perplexity Decider v1.1 has 70.6; the other two are not disclosed. Median latency on their own cards is 84 ms text on one MI325X for Featured3, not disclosed for Perplexity Decider v1.1, 71.4 ms on one GPU for Decision 2.0 Vega-27B and 9 ms on one MI325X for LiquidAI d1-3B.