Opened October 26, Decision Index 0.4.1 Full score for d3 is 67.15; d3-turbo and d3-ultra have their own pages

Xunzhuo Liu posted Decision 3.0 on October 11, 2026, at 08:59 UTC. The attachment is a video, so this page has no still. The post says the weights are open, from 0.6B to 27B, and first on both Jev Decision Index boards in the lab’s evaluation. The collection is Decision 3.0. The repository named on the cards is vllm-project/semantic-router. Decision 2.0 stays on its own page.

The six cards were opened the same day. Each text table says the family’s own row is an internal evaluation, and that the other rows are live board data, text dated 2026-10-10 and vision dated 2026-10-09. d3 prints Decision Index 0.3.1 at 64.7, public 65.5, same-skill 62.8, and new-domain 56.6. The neighbors on that table are Perplexity Decider v1.1 at 62.8, Fastino GLiDE no-thinking at 60.2, Jev at 60.1, Torchcast Decision 27B at 59.9, and Decision 2.0 27B at 55.9. We did not run the weights.

The highlight on the same card prints a different pair. It says the public suite of edition 0.3 is 65.45, all 140,178 public requests answered, and that this is +8.5 over Decision 2.0’s 27B model at 56.97 on the board. The 0.3.1 table’s Decision 2.0 row is index 55.9 and public 57.0. Vega’s 56.5 is a reproduction of edition 0.2.1. Jev at 60.1 on the d3 table is the edition 0.3 reference already printed on the pplx-decider page.

Checkpoint	Parameters on the card	Base named on the card	Index 0.3.1	Public
d3	26.09B, including a 0.46B vision encoder	Qwen3.8-27B	64.7	65.5
d3-flash	8.39B, including a 0.46B vision encoder	Qwen3.5-9B	59.0	57.8
d3-mini	4.54B, including a 0.33B vision encoder	Qwen3.5-4B	54.2	54.9
d3-nano	2.21B, including a 0.33B vision encoder	Qwen3.5-2B	37.2	42.2
d3-lite	0.85B, including a 0.10B vision encoder	Qwen3.5-0.8B	29.0	36.9
d3-edge	0.59B, including a 0.10B vision encoder	Qwen3.5-0.8B	18.0	28.8

d3-flash’s highlight prints public-suite 57.79 for edition 0.3. The four smaller cards’ highlights were not copied here. The collection page’s size chips say 26B, 8B, 5B, 2B, 0.9B, and 0.6B. The parameter column above is the figure each card prints. License on every card opened is Apache-2.0. Each card says it answers choice, yes/no, and score, and that a request can include images and videos.

d3-nano’s own table does not put that checkpoint first. Liquid d1-3B on the same table is index 40.1 and public 49.0, against d3-nano at 37.2 and 42.2. Open d1 prints 48.57 on edition 0.2.1. The September 29 hosted chart is 58.9. The JevBench composite for hosted d1 is 73.0. Those three are different prints from the 40.1.

The d3 vision table, edition 0.3.1, prints 71.6, public 74.2, and private 68.9. The neighbors it prints are Perplexity Decider v1.1 at 70.6, JEV-27B-VL at 69.6 with public 72.8 and private 66.4, and Solomon v1.1 at 66.8. The JEV-27B-VL page does not carry this vision-board row. Perception Test on the d3 card is 77.4 on all 19,140 validation questions, three options, chance 33.3. d3-flash’s vision index is 66.9, and its Perception Test cell is 73.3. Medians on one AMD Instinct MI325X, one request at a time: d3 is 55 ms for text, 273 ms with an image, and 553 ms for a 10-second video. d3-flash is 19.1 ms, 149.4 ms, and 407.6 ms on that same clock.

The cards say text, JSON, images, and videos go in, and that every question in the request sees the images and the frames. Images are read at up to 1.6 megapixels. Videos are read at 2 frames per second, at most 32 frames. The install line asks for transformers==5.17.0. Loading is AutoModel.from_pretrained with trust_remote_code=True, and the method is system_one. No card opened here is deployed by a Hugging Face inference provider.

Opened October 16, 2026 UTC, the same d3 card prints Decision Index 0.3.2 at 66.3, public 65.5, same-skill 62.8, and new-domain 60.5. The highlight now says first on the official text board, and +10.4 over Decision 2.0’s 27B at 55.9. Public 65.5 matches the October 11 public column. New-domain on that earlier table was 56.6. Vision board 0.3.2 prints 71.3, public 74.1, private 68.5. The October 11 vision row was 71.6, 74.2, and 68.9. Perception Test stays 77.4 of 19,140. Text median on one AMD Instinct MI325X is now 54.4 ms. The types listed on the card are still Choice, Yes / No, and Score.

Xunzhuo Liu, 11:35 UTC that day, posted a Box demo running on d3 and d3-flash. The attachment is a video of about 25 seconds, so this page still has no still from that tweet. The text says Jev 1.14 introduced a box type question, so do we, and that a new model is coming. We did not transcribe a score from the video. We did not run the weights.

Opened October 26, 2026, 15:30 UTC, Decision Index 0.4.1 Full score for vLLM-SR d3 is 67.15 on the live space. Public skill stays 65.45. Sibling d3-turbo is 67.31 on that Full column, and d3-ultra is 70.18, second of 229. Date header 06:25:57 GMT October 27, GET https://decision.vllm-sr.ai/v1/models listed vllm-sr/d3 as an alias of vllm-sr/d3-turbo. The October 16 card cell of 66.3 on edition 0.3.2 stays as dated history.