vLLM Decision 3.0: Six Open Models Classify Text, Images, and Video Without Generating Tokens

The vLLM Semantic Router project released Decision 3.0 on October 10, the third generation of its open-weight decision-model family and the first in any open family to accept video as a classification input. The release comprises six models — d3-edge (0.6B), d3-lite (0.8B), d3-nano (2B), d3-mini (4B), d3-flash (9B), and d3 (27B) — all published on HuggingFace under the Apache-2.0 license. Three smaller models have been converted to Apple's Core ML format for on-device inference on Apple Silicon hardware, and all six are now available as hosted endpoints through inference.sh.

Decision Models Answer Questions Rather Than Generate Text

Decision models occupy a distinct role from large language models: rather than producing token sequences, they answer structured questions about an input by returning calibrated probability distributions over a defined option set. The three question types are Choice (select the most appropriate option from a list), Yes-No (binary classification), and Score (a numerical rating within a defined range). Each question receives its own forward pass over the shared input — so a single API call asking three questions about a video clip executes three forward passes, but none of them involves autoregressive decoding, sampling, or beam search. The output is a probability vector, not generated text.

This design makes decision models substantially faster and cheaper per classification than autoregressive language models for tasks like routing, retrieval filtering, and content tagging, where the answer space is bounded and defined in advance. The trade-off is inflexibility: a decision model cannot produce an answer outside its option set or explain its reasoning. For structured automation pipelines, those constraints are typically a feature rather than a limitation.

Decision 3.0 extends this architecture to video inputs, which prior open decision model families did not support. Whether the video representation flows through the same multimodal encoder as image inputs or requires a separate processing path is not detailed in the public model cards; the claim of video support is confirmed by the release documentation.

Benchmark Gains Confirm Progress but a Score Gap Remains

Performance on the Jev Decision Index 0.3.1 — a chance-corrected benchmark across five equal-weight categories (Tools and Automation, Retrieval and Classification, Language Understanding, Knowledge and Reasoning, Arts and Human Judgment) — shows substantial improvement over the prior generation. The d3-lite (0.8B) reached 36.93, up 16.4 points from its Decision 2.0 equivalent. The d3-edge (0.6B) scored 35.91, a gain of 12.1 points. The d3-mini (4B) scored 47.17, up 10.7 points.

Scores for the two largest models — d3-flash (9B) and d3 (27B) — had not been published on the index at the time of this article. The highest published score on the 0.3.1 sub-version currently belongs to Torchcast Decision 27B at 59.9.

The benchmark's dominant closed-model score — TypeSafe Jev's 57.89 — was posted on the older Jev Decision Index 0.2.1, making a direct numerical comparison unreliable. TypeSafe Jev launched in September 2026 with a $40 million seed round led by DCVC and uses a training method its team calls Reinforcement Learning for Calibrated Decisions (RLCD), which it claims improves calibration — meaning the model's reported confidence more closely matches its actual accuracy. This calibration claim has not been independently replicated for open models including Decision 3.0.

Open Weights Close the Deployment Gap That Closed APIs Create

TypeSafe Jev prices at approximately $0.0004 per decision — a number that scales with request volume for production systems. Decision 3.0's Apache-2.0 licensing makes it self-hostable at zero per-decision cost, with hardware and compute as the only ongoing expense. The inference.sh listing provides a hosted alternative with GPU-time billing for teams that prefer managed infrastructure without per-call pricing.

The Core ML conversions extend that deployment flexibility to Apple Silicon, covering d3-lite, d3-nano, and d3-mini. On-device inference eliminates round-trip latency and removes data from cloud infrastructure entirely — relevant for enterprise workflows with privacy constraints, offline requirements, or edge deployment targets where a cloud API is not viable.

The vLLM Semantic Router project also confirmed an AMD partnership focused on inference optimization, signaling continued investment in performance for non-NVIDIA hardware. The project's NeurIPS 2025 paper established its research lineage as a routing framework for Mixture-of-Modality deployments.

Read more: Laya open decision model review: fast but near-chance on unseen tasks

Score Gap and Calibration Claims Remain Unverified

The performance gap between Decision 3.0 and TypeSafe Jev is real, even accounting for benchmark version differences. The d3-mini (4B) scored 47.17 on Jev Decision Index 0.3.1 while Torchcast Decision 27B — a substantially larger open model — reached 59.9, and TypeSafe Jev posted 57.89 on the older 0.2.1. Whether Decision 3.0's 27B model narrows that gap once scored remains unknown.

GPTS24's coverage of Laya, another open decision model, found it running near chance on classification tasks outside its training distribution — a reminder that benchmark scores on a fixed index may not predict generalization to novel domains. Decision 3.0's gains across model sizes suggest stronger generalization than its predecessor, but independent out-of-distribution testing has not yet been reported. TypeSafe Jev's specific RLCD calibration advantage — if it holds outside the benchmark — may matter more for production deployments than headline accuracy scores suggest.

The d3 27B score on the Jev Decision Index 0.3.1, and the first reports of Decision 3.0 performance in real production pipelines, will determine whether the open-weight field has genuinely narrowed the capability gap to TypeSafe Jev or has primarily closed the modality and deployment flexibility gaps while leaving the accuracy gap intact.

Read more: TypeSafe Jev returns decisions without generating text