d3-turbo: vLLM Semantic Router's upgraded 27B decision model reads the input once for all questions

Source · Model card: Decision 3.0 (d3)

publicado (OK do Fabricio 20/10 BRT).

vLLM Semantic Router published d3-turbo on Hugging Face on 18 October 2026 (23:14 BRT) and announced it on X on 19 October (21:42 BRT). It is the upgraded 27B member of Decision 3.0: Apache-2.0, about 26.1B parameters, text, images and video, and box and track questions.

What changes. d3-turbo reads the input once per request and answers every question from that read (“StateOnce”); d3-flash re-reads it for each question. The card’s table, on one AMD Instinct MI325X, shows 302 ms for one question about a photo (d3-flash: 154 ms) and 867 ms for 64 questions (d3-flash: 8.26 s). The X post says “15x faster than d3 flash”; the card’s own table shows 9.5× at 64 questions, with more at 255. The free API at decision.vllm-sr.ai serves d3-turbo, and on that API vllm-sr/d3 names the same model.

Vendor numbers (UnverifiedClaim). The card gives 70.2 on the Decision Index text board and 72.9 on the vision board, both labelled “our estimate”, and 71.7 on the public suite with the official 0.3 kit (d3: 65.5). The X post by Xunzhuo Liu (19 Oct, 21:42 BRT) says “#1 on both Jev Decision Index boards”.

Independent boards (checked 20 Oct).

• Decision Index 0.4, published on 20 October, now computes the Full score from the private parts only (60% same skills, 40% new domains) and keeps 0.3.2 as an archive. It lists d3-turbo at Full 69.04, #6, inside the same tie band as d3 (69.06, #4); mammoth-1 is #1 at 71.24. Its public score is 71.61. So the “#1” claim does not hold on the official text board.
• JevBench open weights (draw v1.6.35): capability 82.6, Intelligence 74.8, calibration 90.4, measured p50 0.144 s on-GPU. Cost $0.22 per 1,000 decisions is 6.75× the Jev reference, outside the board’s cost cap, so its official rank is #93.

Score. The Decision 3.0 page is scored on d3 27B and stays at 7.0 / 7.6 / 11. The page has a d3-turbo section.