# What Does the RTEB Benchmark Actually Measure, and Why Does It Matter?

According to the Hugging Face blog post published July 16, 2026, the RTEB benchmark evaluates retrieval embeddings across three core tasks: single-query retrieval, multi-query retrieval, and agentic retrieval—where an AI agent must iteratively refine its search based on prior results. NVIDIA's Nemotron 3 Embed scored 82.3 overall, surpassing Google's Gecko (79.1) and Cohere's Embed v4 (77.8).

This is significant because most existing benchmarks, like MTEB, only test single-query accuracy. RTEB, introduced in late 2025 by a consortium including Hugging Face and Microsoft, was designed to simulate real-world agentic behavior. NVIDIA's win suggests its model is better at the kind of iterative, context-aware retrieval that powers autonomous AI agents—a rapidly growing enterprise use case.

# How Did NVIDIA Achieve This Lead Over Google and Cohere?

NVIDIA's technical report, linked from the Hugging Face blog, reveals that Nemotron 3 Embed uses a novel two-stage training pipeline. First, it is pre-trained on a curated corpus of 2.1 trillion tokens, with a focus on multi-turn dialogue and search logs. Second, it undergoes contrastive fine-tuning using a synthetic data generator that creates realistic multi-step retrieval scenarios. This approach directly targets the agentic retrieval task that RTEB prioritizes.

In contrast, Google's Gecko, announced in April 2026, relies on a single-stage distillation from a larger LLM, which the RTEB results suggest is less effective for iterative retrieval. Cohere's Embed v4, released in March 2026, focuses on multilingual support but sacrifices agentic performance. According to NVIDIA's blog, the company trained the model on "over 1 million synthetic agentic retrieval trajectories," a scale that competitors have not matched publicly.

# Who Gains and Who Loses From This Benchmark Shift?

The biggest winner is NVIDIA, which now has a concrete benchmark win to sell into enterprise AI stacks. Companies building agentic systems—such as Salesforce's Agentforce and Microsoft's Copilot agents—are natural customers. The biggest loser is Google, whose Gecko model was marketed as the default for retrieval-augmented generation (RAG) but now trails in the agentic benchmark that matters most. Cohere, which has long dominated MTEB, must now decide whether to pivot its embedding strategy or risk being seen as a single-query specialist.

Verdict: NVIDIA Nemotron 3 Embed is the clear leader for agentic retrieval, but Google and Cohere have the resources to catch up within 12 months if they prioritize multi-step training data.

# Does This Mean NVIDIA Will Dominate Enterprise Embeddings?

Not necessarily. Benchmark leadership does not guarantee market adoption. NVIDIA must now build a developer ecosystem around Nemotron 3 Embed, including integrations with popular RAG frameworks like LangChain and LlamaIndex. According to the Hugging Face blog, the model is available under an open-source license on Hugging Face, which lowers the barrier to entry. However, enterprise buyers may still prefer Google's or Cohere's existing cloud integrations and support contracts.

Furthermore, the RTEB benchmark itself may evolve. If Google or Microsoft push for a new version that emphasizes different tasks—such as low-latency retrieval or multilingual support—NVIDIA's lead could shrink. For now, though, the data is clear: NVIDIA has the best model for agentic retrieval, and that matters for the next wave of AI applications.

1. By Q1 2027, Google will release an updated Gecko model that scores above 81 on RTEB, but NVIDIA will maintain the #1 spot with a 2-3 point lead.
2. Cohere will acquire or partner with a synthetic data startup within 6 months to close its agentic retrieval gap.
3. OpenAI will either release a new embedding model by Q4 2026 or exit the standalone embedding market entirely, focusing on end-to-end agent solutions.

• NVIDIA's win is less about raw compute and more about synthetic data strategy—a lesson for all model developers.
• The RTEB benchmark is now the de facto standard for agentic retrieval; companies ignoring it will lose enterprise credibility.
• Enterprise AI buyers should evaluate embedding models on agentic retrieval tasks, not just MTEB scores.
• Google and Cohere have a 12-month window to catch up, but their existing cloud integrations give them a distribution advantage.
• OpenAI's absence from the embedding race is a strategic mistake that will cost it enterprise market share in agentic workflows.