vLLM-SR Releases Decision 3.0, an Open Multimodal Suite That Tops Jev on Text and Vision

The vLLM Semantic Router project has released Decision 3.0, a family of six open-weight multimodal decision models covering a parameter range from 600 million to 27 billion, all under an Apache-2.0 license. The project's reported results place the flagship 27-billion-parameter d3 model at 65.45 on the public Jev Decision Index — the primary benchmark for this model class — above both TypeSafe's closed-API Jev system and Microsoft's recently launched Decision-1. Model cards for all six Decision 3.0 variants are available on the vLLM-SR HuggingFace organization page.

The timing is pointed. Microsoft announced Decision-1 on October 9, 2026, positioning it as an enterprise answer to TypeSafe's Jev in the emerging decision-model category. Two days later, the vLLM-SR project published all six Decision 3.0 variants at no cost, with weights available for download, fine-tuning, or self-hosted serving. TypeSafe and Microsoft have aligned on identical pricing — $0.042 per million input tokens, output tokens free — and Decision 3.0's open weights make that per-call floor irrelevant for teams running their own inference stack.

Decision Models Are Not Language Models: What the Architecture Actually Does

The defining characteristic of a decision model is what it does not do: generate free-form text. Where a large language model produces tokens one at a time and can return anything from a sentence to a function call, a decision model accepts structured application state and a set of typed questions, then returns a probability-weighted answer drawn from a predefined schema. The three output primitives in the Decision 3.0 design are Choice (select one option from a supplied list), Noul (binary yes/no determination), and Score (a numeric rank on a defined scale).

That constraint is the point. An application routing a support ticket, assigning a pull request to a reviewer, or selecting a coverage determination from an insurance policy list does not need a paragraph of explanation — it needs a calibrated probability distribution across its options. TypeSafe has called its approach to probability calibration RLCD (Reinforcement Learning for Calibrated Decisions), which trains the model so that 80% confidence corresponds to empirical accuracy roughly 80% of the time. vLLM-SR has not publicly documented Decision 3.0's post-training methodology, but calibration is implied by its published benchmark performance.

The flagship d3 is built on Qwen3.8-27B, Alibaba's multimodal model supporting text, images, and video as context. The five smaller variants — d3-flash (9B), d3-mini (4B), d3-nano (2B), d3-lite (0.8B), and d3-edge (0.6B) — use Qwen3.5, Alibaba's lightweight model family, and carry the same multimodal input capability through that base. All six models are Apache-2.0.

Read more: Laya, the open rival to TypeSafe's Jev, fast but near-chance on unseen decision tasks

How Decision 3.0 Scores Against Jev and Microsoft-Decision-1

The primary evaluation surface for this category is the Jev Decision Index, a third-party leaderboard hosted on HuggingFace by a community developer known as multimodalart. The index evaluates models against 140,178 requests across 37 benchmarks organized into five equal-weight areas: Tools and Automation, Retrieval and Classification, Language Understanding, Knowledge and Reasoning, and Arts and Human Judgment. Scores run from 0 to 100.

Decision 3.0 reports 65.45 for d3 on the public suite (benchmark version v0.3.1). Microsoft-Decision-1's score of 53.2 on the same index appears in Decision 3.0's release materials; that figure has not been independently confirmed from the HuggingFace leaderboard itself and should be read as the project's reported comparison rather than a neutral third-party finding.

A single developer ran 75 questions about correct WebMCP tool selection across all three systems and found Decision 3.0 outperforming both competitors on that task set. That test has not been independently replicated.

The Jev Decision Index was designed and is maintained in the TypeSafe ecosystem, predating Decision 3.0. Its 37 component tasks and weighting structure were established before the Decision 3.0 models existed, which introduces potential design-side alignment favoring TypeSafe-adjacent architectures. Neither ranking should be treated as a neutral external measurement until independent evaluators publish reproductions.

Read more: TypeSafe's Jev delivers structured AI decisions without generating a single word

Deployment Range: AMD Server to Apple Silicon

vLLM-SR ran latency benchmarks for d3 on AMD Instinct MI325X GPUs under the vLLM inference stack, a hardware choice that reflects vLLM's ongoing investment in AMD ROCm compatibility and establishes AMD as a first-class deployment target for the project. The d3 model card reports median latency of 55 ms for text requests and 273 ms for image requests on a single MI325X, measured one request at a time. Decision models' bounded output structure — a probability distribution over a defined schema rather than open-ended generation — may allow higher per-GPU throughput than equivalent LLM serving workloads.

At the other end of the size ladder, d3-lite (0.8B) was benchmarked running under Apple's Core ML on an M5 Pro Mac, achieving approximately 8.3 pull-request routing decisions per second on a concrete software task: assigning incoming GitHub PRs to code-area reviewers based on file diffs and PR descriptions. A small engineering team generating 30 PRs per working day would exhaust the router's per-second capacity in seconds, leaving the Mac available for other work between bursts.

Where the Benchmark Evidence Falls Short

Several claims in the Decision 3.0 release are not yet independently established. The project's competitive ranking — d3 above Jev, Jev above Microsoft-Decision-1 — comes from the project's own evaluation runs on a benchmark maintained in the TypeSafe ecosystem. Independent researchers have not published reproductions of Decision 3.0's 65.45 score.

The deeper open problem is schema generalization. When GPTS24 covered Laya — an earlier open-weight decision model that also challenged Jev on this benchmark — its performance dropped to near-chance on decision schemas not represented in its training data. That failure mode is architecturally plausible for Decision 3.0 as well, since no decision model project has publicly disclosed its training schema distribution. The benchmark's 37 task types are known in advance by every evaluated model; what happens on genuinely novel schemas outside those 37 types remains an open question across the category.

Open Weights Change the Category's Economic Baseline

Microsoft's road map for Decision-1 includes planned rebasing on MAI and OpenAI models, according to reporting by TestingCatalog, suggesting the company views its current Qwen3.5-9B foundation as temporary. That update may shift benchmark standings when it ships, though no fixed date has been announced.

What changes more immediately is the cost structure. TypeSafe and Microsoft have each set $0.042 per million input tokens as the decision-model API price — a floor that disappears for teams deploying Decision 3.0 on their own hardware. The six-model ladder from d3-edge (0.6B) to d3 (27B) is designed for the full range of deployment constraints, from edge devices to multi-GPU server clusters, without requiring architecture changes as workloads scale. Whether that open-weight calibration holds up against unseen schema types at production volume — the question the Laya analysis left open — is now one the decision-model community has the infrastructure to answer for itself.