The d3-ultra score is the card
The d3-ultra card lists 26.09 billion parameters under Apache-2.0 and says it returns probabilities. The scores come from a private, unpublished test.
The short version
No. The model card says both question sets are unpublished. The parameter count, 26.09 billion including a 0.46 billion vision encoder, and the Apache-2.0 license are on the card.
The card
The Hugging Face card for vllm-sr/d3-ultra describes a multimodal decision model from vLLM Semantic Router. It lists 26.09 billion parameters, including a 0.46 billion vision encoder, inputs of text or JSON plus images and video, and an Apache-2.0 license. It says a request can ask several kinds of question, including a choice, a yes or no, a score, a box on an image or a video moment, and a track across a video. It says the model reads the input once and returns a probability for every answer, without generating text. The context line is 131,072 tokens per question. MarkTechPost wrote the model up on October 26, which is how it reached the wire.
The score is not a public set
The card says the question sets behind its comparison are private, and that the comparison is the team's own test rather than an official leaderboard. MarkTechPost reports a 97.8 score on one SR-Text slice and 54.6 percent on an SR-Box slice, 5.0 points below a sibling called d3-turbo. Those numbers are only as checkable as the hidden questions. Nanbeige's table was filed because a lab table is the lab's. The same rule applies, with a stricter limit: a reader cannot rerun a private set. This page will not treat 97.8 as a ranking against models that sat a different exam.Nanbeige's table.
The sibling recipe is not this card
vLLM's recipe page for vllm-sr/d3, the sibling, says that model was trained from Qwen3.8-27B and describes how vLLM serves a decision checkpoint. The ultra card read here does not name its base checkpoint. Moving the sibling's base onto ultra would invent a line the ultra card omitted. This page also will not copy serving commands, package pins, or a training recipe from either document. Weights a reader can fetch are a different brief: fetchable weights are not the same thing as a public test set.Weights you can fetch.
What the card claims the model is for
The card's examples are routing, classification, and pointing at something in an image or a video, answered as probabilities. A probability is not a generated paragraph, which is the distinction the card draws with ordinary chat models. Terra's addendum, filed beside this one, is a hosted general model with a price and a warning-test rate. d3-ultra is a decision checkpoint with a private scoreboard. They share a desk and not a dependency. TokenGate remains the note on how lab papers time their claims.Terra's warning rate.
Hardware in the card is the team's rig
The card and MarkTechPost describe measurements on one AMD Instinct MI325X with BF16 weights. That is the machine the team used. It is not a minimum-spec certificate, and it is not a claim that the model is fast on every accelerator. A latency range in the card, from a few hundred milliseconds to a couple of seconds as the question count grows, is the team's median on that machine. This page did not rerun it.What a context window bounds.
On-device is a different bound
The on-device inference brief is about what stays on a handset. d3-ultra's card describes a 26-billion-parameter model measured on a data-center accelerator. A reader should not collapse those into one deployment story. The free API the card mentions is the team's hosted demo. A demo key is not a support contract, and this page will not quote its limits as a promise. Asked directly, How big is d3-ultra, on the card? 26.09 billion parameters, including a 0.46 billion vision encoder. The card lists a context of 131,072 tokens per question and an Apache-2.0 license.What on-device inference keeps.
Who the card is for
The card addresses developers who want a probability over a fixed set of answers, including answers about an image, from one open-weight checkpoint. People looking for a chat transcript will not get one from the behavior the card describes. People looking for a leaderboard they can audit will not get one either, until the questions are released. Asked directly, What does the d3-ultra card say the model returns? A probability for each answer. The card says it does not generate text. It reads an input once per request and answers every question on that read.
What this page will not paste
MarkTechPost's piece includes install and serve lines. They are omitted here. A model card's job, for this desk, is the size, the license, the output type, and the status of the test. The recipe's hardware notes about memory and a nightly build belong to operators of the sibling, and they are not copied as steps. AKV's paper was a cache result. It is not a decision model. Asked directly, Are the SR-Text questions public? The card says both question sets are private. MarkTechPost repeats scores from that private test, including 97.8 on one slice and 54.6 percent on another.The AKV paper.
The record
The record is the d3-ultra model card, MarkTechPost's October 26 account of that card, and the sibling recipe only for the line this brief refuses to transfer. Parameters, license, and the private-test warning are the card's. The 97.8 and 54.6 figures stay the team's, on questions the team has not published. The short answer this page will stand on is this: No. The model card says both question sets are unpublished. The parameter count, 26.09 billion including a 0.46 billion vision encoder, and the Apache-2.0 license are on the card.
Sources
The reports this brief is filing. Futureweb did not republish them.
Questions
What does the d3-ultra card say the model returns?
A probability for each answer. The card says it does not generate text. It reads an input once per request and answers every question on that read.
How big is d3-ultra, on the card?
26.09 billion parameters, including a 0.46 billion vision encoder. The card lists a context of 131,072 tokens per question and an Apache-2.0 license.
Are the SR-Text questions public?
The card says both question sets are private. MarkTechPost repeats scores from that private test, including 97.8 on one slice and 54.6 percent on another.
Does the card name a base checkpoint?
The ultra card, as read here, does not. The vLLM recipe for the sibling model d3 says that sibling was trained from Qwen3.8-27B. This page will not move that line onto ultra.
Is a private score comparable to Nanbeige's table?
Nanbeige's table was the lab's own public comparison. d3-ultra's table cannot be rerun from the card, because the questions are withheld.