vLLM Semantic Router releases Decision 3.0: five open decision models, 0.8B to 27B

The vLLM Semantic Router team has released Decision 3.0, a family of five multimodal decision models spanning 0.8B to 27B parameters, all under the Apache-2.0 license. The models read text, JSON and images, then answer typed questions with calibrated probabilities instead of generated text that developers have to parse.

The release matters because routing and guardrail layers sit at the front of nearly every production LLM stack. Teams that currently prompt a general-purpose LLM to classify, route or gate requests can replace brittle output parsing with threshold-able scores from a purpose-built model.

What is Decision 3.0?

Decision 3.0 is a set of open decision models from vLLM Semantic Router that return probabilities instead of text. A developer passes a state (text or JSON), optional images, and named questions. Each question takes one of three forms: a Choice, a Yes/No, or a Score on a scale. The model answers all questions in a single call and returns a probability for every answer.

This is what the team describes as the 'system one' pattern for routers and guardrails. Instead of prompting an LLM and parsing its output, callers get calibrated scores they can threshold.

The five sizes are:

• d3-lite — 0.85B parameters
• d3-nano — 2.21B
• d3-mini — 4.54B
• d3-flash — 8.39B
• d3 — 26.09B

Context length is not disclosed. The latency figures the team published were measured on a single AMD Instinct MI325X GPU, with BF16 weights and no official quantized variants listed.

How does d3 work?

Each model is a fine-tune of a Qwen base. The 27B d3 builds on Qwen3.8-27B; the four smaller models build 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, and every question sees all images.

Each question gets its own forward pass over the input. The install targets transformers==5.17.0, the optional flash-linear-attention package speeds up the linear-attention layers, and loading requires trust_remote_code=True.

How does Decision 3.0 perform on benchmarks?

Through its model cards, the research team cites the Jev Decision Index 0.3.1, a board that ranks open reproductions of TypeSafe's Jev decision system. All d3 numbers come from the team's internal evaluation rather than an independent run.

On the text board, d3 scores 64.1, ahead of Perplexity Decider v1.1 at 62.8 and Jev at 60.1. One caveat: Torchcast Decision 27B posts a higher public-suite score, 65.1 versus d3's 64.9. On the vision board, d3 scores 71.6 against 70.6 for Perplexity Decider v1.1.

The size-efficiency story is the strongest part of the release. d3-flash, at 9B parameters, 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 at the 27B size to +15.1 at 0.8B.

Class leadership is not uniform across the lineup, though:

• d3-mini (4B) and d3-flash (9B) lead both their text and vision tables.
• d3-lite (0.8B) leads on text but ranks third on vision, scoring 41.1 behind JPT-0.8B at 42.1.
• d3-nano (2B) 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. Its best single result is 97.7 on KIE (CORD+FUNSD) document extraction. Its worst is 34.1 on R-Bench-M. It is also weaker on MMMU-Pro vision (46.2) and Hateful Memes moderation (44.6).

The self-reported nature of the scores is the release's main caveat: every d3 number is the team's internal evaluation scored against live board data, not an independent audit.

How fast are the models?

The team published median latencies for one request at a time on one AMD Instinct MI325X, per each model card:

• d3-nano: 21.7 ms text, 92.0 ms with an image
• d3-lite: 23.5 ms text, 60.7 ms with an image
• d3-mini: 28.5 ms text, 115.0 ms with an image
• d3-flash: 29.7 ms text, 171.8 ms with an image
• d3: 84 ms text, 342 ms with an image

All models were trained on AMD Instinct MI325X GPUs. The sub-30 ms text latencies across four of the five sizes put the smaller models in range for inline request gating.

How does d3 compare with other decision models?

The 27B d3 faces several open competitors. Perplexity Decider v1.1 shares the same Qwen3.8-27B base and Apache-2.0 license, accepts text and images, and discloses an 8,192-token context per decision. Decision 2.0 Vega-27B is a 29.37B-parameter model on the same base, handles text and JSON only, and lists a 32,768-token context. LiquidAI's d1-3B is a 3.12B model built on LFM2.5-VL-3B under the LFM 1.0 license, also with a 32,768-token context.

On the Decision Index, the gap is wide: per the d3 cards' readings of the 0.3.1 board, d3 scores 64.1, Perplexity Decider v1.1 62.8, Decision 2.0 Vega-27B 55.9, and LiquidAI d1-3B 40.1. On latency, LiquidAI d1-3B claims 9 ms on one MI325X and Decision 2.0 Vega-27B lists 71.4 ms on one GPU, while Perplexity does not disclose latency at all.

Two gaps stand out against peers: d3 does not disclose its context length, where competitors publish 8,192 or 32,768 tokens, and its vision capability is matched by LiquidAI and exceeded in claimed speed.

What should developers take away?

The headline facts are straightforward. Five Apache-2.0 decision models shipped, from 0.8B to 27B, all with image input. The flagship d3 tops its text (64.1) and vision (71.6) tables on internal evals. The 9B d3-flash beats the previous generation's 27B model outright. Median text latency runs from 21.7 ms to 84 ms on a single MI325X.

The caveats are equally concrete. The 0.8B and 2B models do not lead every board. Scores are self-reported against live board data. Context length remains undisclosed.

For teams building routing or guardrail layers, d3-flash looks like the practical center of the lineup: it beats the prior 27B generation at a third of the size and answers text requests in under 30 ms. Whether the internal numbers hold up under independent evaluation on the Jev Decision Index will determine if Decision 3.0 keeps the top of the board.