d3-ultra 27B multimodal decision model scores 87.2

The vLLM Semantic Router team has released d3-ultra, a 26.09 billion parameter multimodal decision model under Apache-2.0 license. The model operates as a classification and routing tool for developers building agent systems and production applications. Unlike text-generating language models, d3-ultra returns probability scores for predefined answer choices rather than generating new text.

What changed

d3-ultra is the latest member of the Decision 3.0 family, which ranges from 0.6B to 27B parameters. The model processes input once per request and returns probabilities across all answer options. This single-read approach enables routing, classification, visual grounding, and video analysis from a single API call.

The architecture consists of 26.09 billion parameters in a dense configuration, including a 0.46B vision encoder. It supports 131,072 tokens of context per question and uses a Qwen 3.5 backbone with 64 layers split between 48 linear-attention layers and 16 full-attention layers. Current testing relies on BF16 weights on AMD Instinct MI325X hardware, though the team has not yet released an official quantized variant.

The model handles five question types: Choice (select one option), Yes/No (return affirmative probability), Score (rate on a scale), Box (locate regions in images or video), and Track (follow targets across video frames). Questions can incorporate images within them or as answer options.

Deployment paths include the Transformers library with flash-linear-attention support, the vLLM Semantic Router Engine integrated in October 2026, a free API endpoint, and a Hugging Face demo space. Developers can self-host the model or use hosted endpoints based on application requirements.

Why it matters

Decision models occupy a distinct niche in the AI ecosystem. Rather than generating text, they answer questions with probability scores, which reduces hallucination risk and improves predictability for deterministic tasks. This architecture suits classification scenarios like ticket routing, SQL validation, content moderation, and visual grounding where probabilistic answers matter more than generated content.

The multimodal capability addresses a real workflow problem: teams can now handle images and video alongside text without managing multiple specialized models. A single model becomes responsible for text queries, image analysis, and video understanding, simplifying deployment and reducing operational complexity.

The Apache-2.0 license and open availability lower adoption barriers compared to proprietary alternatives. Organizations gain the ability to self-host without vendor lock-in, addressing data residency and security concerns that prevent some teams from using closed models. This flexibility particularly appeals to enterprises with strict infrastructure requirements.

Performance metrics matter for production systems. On text tasks, d3-ultra ranks highest among its family on self-reported benchmarks. Latency characteristics also matter: approximately 304 milliseconds for a single question about a photo, with additional questions adding roughly 7.3 milliseconds each. These characteristics help teams understand whether the model fits their response-time requirements.

What to test

Before deploying d3-ultra, teams should verify several technical and performance claims:

• Benchmark results are self-reported. The team tested d3-ultra on private SR-Text and SR-Box test sets, not independent third-party benchmarks. The 87.2 score on text tasks needs validation against your specific domain before production use. Public benchmarks may reveal different performance characteristics.

• Box grounding performance lags on natural images. On SR-Box testing, d3-ultra scores 58.2% compared to sibling d3-turbo at 60.0%. The team notes d3-turbo remains stronger on natural photographs while d3-ultra leads on charts. If visual grounding of objects in real-world images is critical, the performance gap warrants evaluation.

• Hardware costs require calculation. AMD Instinct MI325X hardware pricing and availability vary significantly by region and deployment model. Test latency and throughput on your target hardware to confirm the 2.16 second benchmark applies to your infrastructure.

• No quantized version exists yet. Current weights require BF16 precision hardware, limiting deployment options. When quantized variants arrive, measure quality loss before switching production systems, as decision models may degrade differently under quantization than text-generation models.

• Context limits are hard boundaries. Questions exceeding 131,072 tokens receive max_length_exceeded errors rather than silent truncation. This behavior differs from some models and requires careful prompt engineering.

• Performance varies significantly by task category. d3-ultra achieves 97.8% on tools and automation tasks but only 54.6% on box grounding across nine domains. Identify which task categories matter for your use case and test performance there specifically.

The conclusion

d3-ultra represents a practical option for teams building classification and routing layers within agent architectures. The 87.2 score on text tasks and Apache-2.0 licensing reduce barriers compared to closed alternatives. However, the performance data derives from private benchmarks without independent verification. The box-grounding weakness on natural images and the absence of quantized weights warrant careful evaluation for specific use cases.

Teams should conduct domain-specific testing before committing to production deployments. The open availability and multiple deployment options reduce switching risk, but benchmark verification remains important. Watch for quantized releases and community-published benchmark results as the Decision 3.0 family develops further.

Related reading

• Drex for Search scores 92.05% on BrowseComp-Plus
• Nanbeige4.3-33B scores 80.9 on Terminal-Bench
• Carbon-X: 14.8B open DNA model scores 66.7% on benchmarks