Published
Updated
Opened October 26, Decision Index 0.4.1 Full score for d3-ultra is 70.18, second of 229
Xunzhuo Liu posted d3-ultra on October 27, 2026, at 03:01 UTC. Opened October 26, Decision Index 0.4.1 Full score is 70.18, second of 229, +3.03 over d3 at 67.15. The card's private SR-Text mean is 87.2. Apache-2.0. This desk did not run the weights.
Xunzhuo Liu posted d3-ultra on October 27, 2026, at 03:01 UTC. The attachment is a video of about 65 seconds. The post names an open 27B multimodal decision model, number two of 230 on the Jev Decision Index text board, +3.03 over d3, with Box and video track, and links vllm-sr/d3-ultra. vLLM quoted that post at 03:23 UTC. Marktechpost, 05:52 UTC, printed the card’s SR-Text mean and the 255-question clock. The share image on this page is that SR-Text and SR-Box chart. The earlier Decision 3.0 family is on Decision 3.0. Sibling d3-turbo has its own page.
The card, opened October 27, prints 26.09B parameters including a 0.46B vision encoder, Choice, Yes / No, Score, Box, and Track, and 131,072 tokens per question. License is Apache-2.0. The API model list names the base as Qwen3.8-27B. The private SR-Text table prints d3-ultra 87.2, GLiDE V1.2 85.1, Liquid d1-ultra 83.4, d3-turbo 83.1, d3 81.4, and Jev 1.14 73.1. SR-Box on ten domains prints d3-ultra 58.2% and d3-turbo 60.0%. GLiDE, d1-ultra, and d3 have no Box score on that table. Both suites are private. On one AMD Instinct MI325X, one question about a photo is 304 ms and 255 questions are 2.16 s. The free host is decision.vllm-sr.ai. The collection is Decision 3.0. The repository named on the family cards is vllm-project/semantic-router.
Opened October 26, 2026, 15:30 UTC, data/index.json is Decision Index 0.4.1, 229 models, generated 2026-10-26T15:30:58Z. Full score for vLLM-SR d3-ultra is 70.18, second of 229. Fastino GLiDE 1.2 is first at 71.52. d3 on that file is 67.15. 70.18 − 67.15 = 3.03, the gap in the launch post. Public skill for d3-ultra is 71.57. Jev 1.14 on the same file is a reference at 60.24. The vision board data/vision.json, generated 2026-10-26T15:31:01Z, edition 0.4, 56 entrants, prints d3-ultra full 70.59, public 75.46, private 65.73, rank 11. d3 on that vision file is 71.26. We did not rerun either board.
Date header 06:25:57 GMT October 27, GET https://decision.vllm-sr.ai/v1/models returned HTTP 200 and listed vllm-sr/d3-ultra. The same list prints vllm-sr/d3 as an alias of vllm-sr/d3-turbo. Date header 06:25:58 GMT, POST /v1/systemone without a key returned HTTP 401, missing_api_key. This desk has no key for that host. We did not run the weights.
This site's reading
Editorial notes evaluating claims against primary sources, contextualizing findings alongside related implementations, and defining technical terms.
Verify
Primary post is @XunzhuoLiu, October 27, 2026, 03:01 UTC, status 2114915380926017591. The text names d3-ultra as an open 27B multimodal decision model, number two of 230 on the Jev Decision Index text board, +3.03 over d3, with Box and video track, and links huggingface.co/vllm-sr/d3-ultra. The attachment is a video of about 65 seconds. The share image on this page is the card's SR-Text and SR-Box chart. @vllm_project, 03:23 UTC, status 2114920881525051889, quoted that post. @Marktechpost, 05:52 UTC, status 2114958464770539742, prints the team's SR-Text mean 87.2 against GLiDE V1.2 at 85.1 and d1-ultra at 83.4, and 255 questions about one photo at 2.16 s on one AMD Instinct MI325X.
The Hugging Face card vllm-sr/d3-ultra, opened October 27. Parameters 26.09B including a 0.46B vision encoder. Decision types Choice, Yes / No, Score, Box, and Track. Context 131,072 tokens per question. License Apache-2.0. Base named on the API metadata is Qwen3.8-27B. The card's private SR-Text table prints d3-ultra 87.2, GLiDE V1.2 85.1, d1-ultra 83.4, d3-turbo 83.1, d3 81.4, and Jev 1.14 73.1. SR-Box on ten domains prints d3-ultra 58.2% and d3-turbo 60.0%. Both question sets are private. Serving table on one AMD Instinct MI325X: 1 question about a photo 304 ms, 255 questions 2.16 s. Free API decision.vllm-sr.ai. Collection Decision 3.0. The earlier family page is Decision 3.0. Sibling d3-turbo is on its own page.
Opened October 26, 2026, 15:30 UTC, the Decision Index space data/index.json is edition 0.4.1, 229 models, generated 2026-10-26T15:30:58Z. Full score (balanced_skill) for vLLM-SR d3-ultra is 70.18, second of 229. Fastino GLiDE 1.2 is first at 71.52. d3 on that file is 67.15. 70.18 minus 67.15 is 3.03. Public skill for d3-ultra is 71.57. Jev 1.14 on the same file is a reference at 60.24. The vision board data/vision.json, generated 2026-10-26T15:31:01Z, edition 0.4, 56 entrants, prints d3-ultra full 70.59, public 75.46, private 65.73, rank 11. d3 on that vision file is 71.26. We did not rerun the index.
Date header 06:25:57 GMT October 27, GET https://decision.vllm-sr.ai/v1/models HTTP 200. The list includes vllm-sr/d3-ultra, and vllm-sr/d3 as an alias of vllm-sr/d3-turbo. Date header 06:25:58 GMT, POST /v1/systemone without a key returned HTTP 401, missing_api_key. No DECISION_API_KEY in this desk's environment. We did not run the weights.
Compare
70.18 is Full score on Decision Index 0.4.1. 66.3 is the October 16 reading of the d3 card on edition 0.3.2. 87.2 is the card's private SR-Text mean. 58.2% is SR-Box on ten domains. Do not merge 70.18 with 87.2, with 58.2%, or with d3-turbo's 67.31 on the same 0.4.1 file.
71.52 is GLiDE 1.2 on that same Full column. 60.24 is Jev 1.14 on that file. Vision 70.59 is edition 0.4, a different board from the text Full score.
Terms
- d3-ultra
- vLLM Semantic Router's 26.09B Decision 3.0 checkpoint, Apache-2.0, from Qwen3.8-27B. Opened October 26, 2026, Decision Index 0.4.1 Full score is 70.18, second of 229. The card's private SR-Text mean is 87.2.
- SR-Text
- The d3-ultra card's private text suite: 10,411 requests and 73,495 questions. The printed mean of six chance-corrected parts is 87.2 for d3-ultra.
- SR-Box
- The d3-ultra card's private Box suite: 5,678 questions on 4,984 images. Ten-domain mean for d3-ultra is 58.2%. d3-turbo on that table is 60.0%.