d3-ultra

Open-weight 26B multimodal decision model from vLLM Semantic Router that returns answer probabilities instead of generated text.

vLLM Project · Open-weight/community models · model

Published in Feed

Overview

d3-ultra is an Apache-2.0 open-weight multimodal decision model from the vLLM Semantic Router team and the largest member of the Decision 3.0 family, which spans 0.6B to 27B. It reads a text, JSON, image or video input once and returns a probability for each answer to typed questions (Choice, Yes/No, Score, Box and Track) instead of generating text. It has 26.09B dense parameters including a 0.46B vision encoder and takes up to 131,072 tokens per question; inputs over the limit return an error rather than being truncated. The authors report a mean score of 87.2 on their private SR-Text benchmark and a 304 ms p50 for one image question on a single AMD Instinct MI325X.

This is a sourced directory entry. Features and capabilities are reported in the linked sources; this entry does not document independent Aymorphic testing.

Key features

  • Outputs per-answer probabilities in one forward pass, no text generation
  • Five question types: Choice, Yes/No, Score, Box (locate in image or video) and Track (follow a target through video)
  • 26.09B dense parameters with a 0.46B vision encoder; 131,072-token context per question
  • Self-reported SR-Text mean 87.2 (GLiDE V1.2 85.1, d1-ultra 83.4, d3-turbo 83.1) and SR-Box 58.2% versus d3-turbo's 60.0%
  • Runs via Transformers 5.17.0, `vllm-sr serve -e vllm-sr/d3-ultra` (engine support added in PR #5449 on 2026-10-25) or a free hosted API at decision.vllm-sr.ai

Use cases

  • Routing or gating requests in an LLM gateway with a single low-latency classification pass
  • Content and agent-safety judging, where the model card reports 0.99 AUROC on safeagents-judge-bench
  • Visual checks that need a located answer, such as asking where an object appears in an image or video moment

Pricing

Free: Apache-2.0 weights for self-hosting plus a free hosted API key at decision.vllm-sr.ai.

Pricing reflects the source record and may change. Confirm current terms with the provider.

Dates and sources

Released
Not listed
Record updated
3 Nov 2026
Record last checked
3 Nov 2026

Explore the Feed

Other entries filed under open-weight/community models; inclusion does not imply equivalent features.

  • Muse Spark Open

    Meta's open-weight 816B-parameter MoE model (42B active, 1M context) for coding and long-horizon agent work, released Oct 15, 2026.

  • Nemotron 3.5 Bolt

    NVIDIA's open 74B MoE model (9B active) packed into 24GB for local agents on RTX Spark; released Oct 15, 2026.

  • Jais 3

    Six open Arabic-focused language models from 900M to 375B parameters with 512K context, plus Jais Speech for Arabic recognition.

Browse the Feed