Addis PulseStudio

Decision 3.0: A 0.6B-to-27B multimodal decision family, Apache-2.0, on a Qwen base

The vLLM Semantic Router team released a family of small-to-large multimodal decision models that can gate video output locally. The license is clean, the base is familiar, and the one benchmark claim is unverifiable from what was published.

4 min read889 words

What happened

The vLLM Semantic Router team released Decision 3.0, a family of multimodal decision-classification models spanning 0.6B to 27B parameters that accept text, images, and video as input. The models are Apache-2.0 licensed and built on Qwen3.5 and Qwen3.8-27B bases, with a vision encoder folded into each stated parameter count.

Context

General-purpose LLMs can do classification, but they are slow, expensive to run, and return free-text that you have to parse. A purpose-built "decide yes or no, pick A over B" model removes that overhead. The vLLM Semantic Router team has been building routing and decision logic for serving infrastructure; Decision 3.0 is their move into the model layer itself. The Qwen base matters: this is a fine-tune of a widely deployed architecture, not a from-scratch design, which means the community already has serving stacks, LoRA workflows, and text-encoder tooling for it. The announcement landed on 2026-10-12 via r/LocalLLaMA.

How it works

Each model wraps a Qwen language backbone — 0.8B, 2B, 4B, or 9B from the Qwen3.5 line, or the 27B Qwen3.8 variant — with a vision encoder. In the 4B model, 0.33B of the parameters belong to that encoder; in the 27B, 0.46B. Input can be text, an image, or a short video clip. Output is constrained to one of three decision types: choice (pick from a set of options), yes/no, and a third modality the post truncates at "scor," almost certainly scoring, though the full name and semantics are not stated in the excerpt.

Because the release comes from the vLLM Semantic Router team, the implied serving path is a vLLM instance or a local inference server, not a hosted API. No source in the brief provides minimum VRAM per model size, inference latency, or throughput figures, so sizing a deployment requires benchmarking on your own hardware. Training data provenance is not documented in any source.

Our read

The "SOTA decision model family" framing is the team's own descriptor, and the only benchmark cited is the "Jev Decision Index," which no source in the excerpt defines, links, or attributes to a publisher. The #1 claim is self-reported with no methodology visible. That does not invalidate the models; it means the claim is not verifiable from what was published.

What is actually interesting is the shape of the release. A 45× parameter spread, all multimodal, all Apache-2.0, all on a base the community already has serving tooling for. A small studio can run the 4B variant on a single consumer GPU as a local quality gate next to a diffusion pipeline, then scale to 27B when it needs more nuance, without changing the licensing posture at either end. The Qwen base means existing text-encoder nodes and LoRA hooks in a ComfyUI stack may transfer to fine-tuning d3 for a domain-specific decision.

The gap is the third decision type. "Choice" and "yes/no" cover routing and pass/fail. If the truncated "scor" type is continuous scoring, it is the one that replaces a human reviewing frames. But the post does not define the output range, the calibration, or whether scores are comparable across model sizes. Until that is documented, a studio should not wire it into an automated pipeline on faith.

The 0.6B figure in the post title does not match any listed Qwen base (smallest is 0.8B). It may refer to language-only parameters with the vision encoder excluded, but the post does not say so. For someone sizing VRAM, that distinction matters.

What this changes

On Monday, the practical move is narrow: spin up the 4B or 9B d3 model as a separate inference service and point a generated frame or a 2-to-4-second clip at it for a yes/no pass-fail or a choice between two candidate generations. That is a quality-gate step that previously meant a human eyeballing output or a cloud API call. The ComfyUI stack itself does not change; d3 sits alongside it as another local process.

What does not change yet: the scoring modality is underspecified, the benchmark claim is unverifiable from the excerpt, and no source states minimum VRAM per model size. A studio should pilot the yes/no and choice paths before trusting the score path, and should benchmark VRAM and latency on its own GPU before committing to a model size.

License

Apache-2.0. Commercial use is permitted with no usage fees, no revenue ceiling, and no obligation beyond retaining the standard Apache notice in any redistributed code or model weights. A studio can ship a product built on Decision 3.0 without contacting the team or paying a royalty.

Key takeaways

  • Decision 3.0 spans 0.6B–27B parameters, all multimodal (text, image, video), all Apache-2.0, built on Qwen3.5 and Qwen3.8 bases.
  • The #1 benchmark claim is self-reported on the "Jev Decision Index," an index that no source in the excerpt defines, links, or attributes to a publisher.
  • The third decision type (likely scoring) is truncated in the source; output format, range, and calibration are unknown.
  • The 0.6B figure in the title does not correspond to any listed Qwen base (smallest is 0.8B); the discrepancy is unexplained.
  • No source states minimum VRAM, inference speed, training data provenance, or a weight-download URL.

Sources

  1. Decision 3.0: open, multimodal decision models from 0.6B to 27B. Ranks #1 on both Jev Decision Index boards — tier 3
multimodalqwenllmopen-source

How this post was made

Drafted from clustered primary sources by the models below, then read, edited and approved by a human before it was published. The sources are listed in full at the end of the article.

Drafted
Independent sources
1
cluster pair
clef:27b
cluster label
gemma4:12b
radar brief
gemma4:12b
research brief
qwen3.8:27b
draft article
qwen3.8:27b
short script
qwen3.8:27b
seo pack
gemma4:12b
Run
editorial-20261012T135816Z