vLLM Semantic Router 发布 d3-ultra：27B 开源多模态决策模型
vLLM Semantic Router Releases d3-ultra: A 27B Open Multimodal Decision Model - MarkTechPost
vLLM Semantic Router 团队发布 d3-ultra，一个 26.09B 参数、Apache-2.0 许可的多模态决策模型，不生成文本，只对输入一次性读取后按问题类型返回概率。它支持 131,072 token 上下文，在团队自建的 SR-Text 测试中得分 87.2，高于 GLiDE V1.2 的 85.1；但在 SR-Box 定位任务上 58.2% 低于同系列 d3-turbo 的 60.0%。在单张 AMD Instinct MI325X 上，一张图片的 1 个问题耗时 304 ms，255 个问题耗时 2.16 s。所有基准数字均为团队在私有测试集上自报。
正文
Add as a preferredsource on Google
The vLLM Semantic Router team has released d3-ultra, a 27B multimodal decision model under Apache-2.0. The d3-ultra decision model does not generate text. It reads an input once and returns a probability for every answer to every question. For developers, that means routing, classification, visual grounding and video checks from 1 open-weight call.
TL;DR
Size: 26.09B parameters (dense, including a 0.46B vision encoder); 131,072 tokens of context per question.
Runs on: tested on 1 AMD Instinct MI325X with BF16 weights; no official quantized variant is listed yet.
Performance: highest SR-Text score of the 6 decision models in the team’s own private test.
Best: 97.8 on SR-Text tools and automation (chance-corrected accuracy).
Worst: 54.6% on SR-Box (9 domains), 5.0 points below its sibling d3-turbo.
What is d3-ultra?
d3-ultra is a System One decision model from vLLM Semantic Router that answers typed questions with probabilities. It is the newest member of Decision 3.0, a family that spans 0.6B to 27B.
TEXTTABLELAYOUTIs your PDF parser actually getting it right?omni_parse_bench · scores text, tables & layout · open sourceStar on GitHub→Sponsored
You send a state (text or JSON, plus images or videos) and a set of questions. Each question has a type:
Choice: pick 1 of several options.
Yes / No: return the probability of yes.
Score: rate on a scale.
Box: locate something in an image or a video moment.
Track: follow a target across a whole video.
Images can also sit inside questions and options. A quiz can show a photo, then offer 4 etchings as answers.
How does d3-ultra work?
The core idea is state once. The model reads the input once per request and answers every question from that single read.
The config shows a Qwen3.5-architecture backbone with 64 layers. 48 are linear-attention layers and 16 are full-attention layers. The card does not name its base checkpoint. vLLM’s recipe says the sibling d3 was trained from Qwen3.8-27B.
Box answers always return exactly 1 box with a confidence. A confidence near 0 means the target is not in the picture. Questions over 131,072 tokens fail with max_length_exceeded; nothing is silently truncated.
How does d3-ultra perform on benchmarks?
The research team compares d3-ultra with GLiDE V1.2, d1-ultra, d3-turbo, d3 and Jev 1.14 on 2 in-house tests.
SR-Text has 10,411 text requests with 73,495 questions. Most recreate skills from 34 public benchmarks with new items. Another 4,896 questions come from 43 new task families, like SQL checks and ticket routing. Scores are chance-corrected: 0 is chance, 100 is perfect.
d3-ultra leads every area except arts and human taste. There, d1-ultra scores 90.8 against 89.6, within the margin of error.
SR-Box has 5,678 Box questions on 4,984 images across 10 domains. A box counts as correct at IoU 0.5 or higher. d3-ultra scores 58.2% over 10 domains and 54.6% over 9. d3-turbo scores 60.0% and 59.6%. Jev 1.14 scores 45.7% and 46.9%. The team says d3-turbo stays stronger on natural images, while d3-ultra leads on charts.
Latency: on 1 AMD Instinct MI325X with AutoModel, 1 question about a photo takes 304 ms. 255 questions take 2.16 s, about 7.3 ms per extra question.
How do you deploy d3-ultra?
There are 4 routes:
Transformers: AutoModel.from_pretrained("vllm-sr/d3-ultra", trust_remote_code=True) with transformers==5.17.0. The optional flash-linear-attention package adds fast GPU kernels.
vLLM Semantic Router Engine: vllm-sr serve -e vllm-sr/d3-ultra, added in PR #5449 on October 25, 2026.
Free API: the card says decision.vllm-sr.ai serves d3-ultra, with a free key.
Demo: the official Hugging Face Space.
d3-ultra vs GLiDE V1.2, d1-ultra, Jev 1.14 and d3-turbo
Featured3-ultrad3-turboGLiDE V1.2d1-ultraJev 1.14
Parameters26.09B26.09B26B Not disclosedNot disclosed
Context per question131,07232,768 default; 131,072 with long_context40,000 default32,768 262,144
Weights / licenseOpen, Apache-2.0Open, Apache-2.0Open, Fastino Community LicenseHosted onlyHosted only, proprietary
InputsText, JSON, images, videoText, JSON, images, videoText, JSON, imagesText, imagesText, images
Box / trackBox + trackBox + trackNoNoBox only
SR-Text (self-reported by vLLM-SR)87.283.185.183.473.1
SR-Box, 10 domains58.2%60.0%Not supportedNot supported45.7%
Public index scoreSubmitted, not yet on board70.2 Jev DI (own estimate)71.5 Decision Index 0.4.165 AA Decision Index 57 AA Decision Index
PriceFree API, self-hostFree API, self-host$0.04 / M input $0.042 / M input$0.042 / M input
Key Takeaways
d3-ultra is a 27B Apache-2.0 decision model that never generates text.
It reads input once: 255 questions about 1 photo take 2.16 s.
It tops the team’s SR-Text test at 87.2, ahead of GLiDE V1.2 at 85.1.
It trails d3-turbo on Box grounding: 58.2% vs 60.0%.
All benchmark numbers are self-reported on private test sets.
Asif RazzaqWebsite | + postsBio ⮌
Asif RazzaqNace AI Releases Drex for Search for AI Agents: 92.05% on BrowseComp-Plus With 12.6 Search Calls
Asif RazzaqNanbeige Releases Nanbeige4.3-33B, a 33B Open Agentic Coding Model
$uv pip install omni-parse-bench✓text✓tables✓layout→ scored per pageStar on GitHub→Sponsored
RELATED ARTICLESMORE FROM AUTHOR
OpenAI Launches GPT-6.1 Terra, Halving GPT-6.1 Sol’s API Price for Agent Work
Nace AI Releases Drex for Search for AI Agents: 92.05% on BrowseComp-Plus With 12.6 Search Calls
Nanbeige Releases Nanbeige4.3-33B, a 33B Open Agentic Coding Model