vLLM Semantic Router releases 27B open multimodal decision model d3-ultra
Unlike generative models, d3-ultra does not produce text but assigns probabilities to answers across multiple questions in a single input. It supports text, JSON, images, and videos, with a context window of 131,072 tokens per question and a 0.46B-parameter vision encoder included in its 26.09B total parameters.
Key points
- vLLM Semantic Router released d3-ultra, a 27B open multimodal decision model under Apache-2.0
- Supports text, JSON, images, and videos with a 131,072-token context window and 0.46B vision encoder
- Scores 87.2 on SR-Text benchmark, but trails d3-turbo in visual grounding (58.2% vs 60.0%)
The model excels in text-based decision tasks, scoring 87.2 on the team’s SR-Text benchmark (chance-corrected accuracy), surpassing competitors like GLiDE V1.2 (85.1) and its sibling d3-turbo (83.1). However, it underperforms in visual grounding (58.2% on SR-Box), trailing d3-turbo (60.0%). Deployment options include Hugging Face Transformers, the vLLM Semantic Router Engine, a free API at decision.vllm-sr.ai, and a demo on Hugging Face Space. The team reports 304 ms latency for a single photo question on an AMD Instinct MI325X GPU, with 2.16 seconds for 255 questions—7.3 ms per additional question—using BF16 weights. No quantized variants are listed yet.
Model pages: d1-ultra → · Jev 1.14 →
The story so far
2 episodes →- vLLM Semantic Router releases 27B open multimodal decision model d3-ultrathis story
vLLM Semantic Router Releases d3-ultra: A 27B Open Multimodal Decision Model
MarkTechPost · 27 October 2026
Loading the full article…
This text was published by MarkTechPost and written by Asif Razzaq. It is reproduced here with attribution so you can read it in full; the rights remain with the publisher. Read it at the source ↗
The headline, key points and digest above were generated by Digest AI's editorial model from the linked sources. Automated summaries can contain errors: the sources are the record. Spotted a mistake? Tell us. Published by Martin K., who runs Digest AI and handles corrections.
More in Generative AI & Models
All →- Moonshot AI announces Kimi K3.1 Pro with 1M-token context and 3D model creation · 1 src
- OpenAI releases GPT-6.1 Terra with halved API pricing and improved computer use performance · 13 src
- Anthropic reports 3.6x jump in token volume after Claude Opus 5.5 launch · 4 src
- Mirendil announces Fern v0.1 with 10M-token context window · 1 src
- Moonshot AI releases Kimi K3.1 Pro, a 2.8T open-weight model closing gap with proprietary leaders · 24 src
Comments
via GitHub Discussions