Box Integrates NVIDIA Nemotron 3 Embed: 4x Faster Retrieval, 20% Fewer Errors

Executive Overview

At its developer conference on Thursday, Box announced Box Content Retrieval, a new feature built on NVIDIA's open Nemotron 3 Embed models. The system delivers responses 4x faster, reduces errors by 20%, and cuts token consumption by 90%, according to Kari Briski, NVIDIA's VP of Generative AI. The feature will be available to all Box Enterprise Advance customers early next year, while the underlying models are available immediately on Hugging Face, NVIDIA NIM, and vLLM.

Deep-Dive Features & Architecture

Box Content Retrieval is powered by Nemotron 3 Embed models that convert documents into semantic representations capturing concepts rather than just keywords. The family includes three models: Nemotron-3-Embed-8B-BF16, the flagship for high-accuracy critical retrieval; Nemotron-3-Embed-1B-BF16 for cost- and latency-sensitive production use; and Nemotron-3-Embed-1B-NVFP4, optimized for Blackwell architecture for maximum throughput. According to Kari Briski, this architecture translates to 4x faster responses, 20% fewer errors, and 90% lower token consumption, reducing the cost of running enterprise agents.

The model offers open weights, data, and recipes that allow teams to inspect, fine-tune, and deploy on their own infrastructure, with a 32k token context window that reduces truncation in long documents. The NVFP4 path provides 4-bit quantization with quantization-aware distillation (QAD) training to recover accuracy in long sequences, and up to 2x higher throughput compared to BF16 on Blackwell. Aaron Levie, CEO of Box, described the system as integrating five technologies into a single API, relying on Box's metadata and permission controls that are difficult to replicate.

Benchmark & Competitive Performance

On the RTEB benchmark, Nemotron-3-Embed-8B-BF16 leads with a score of 78.5% and achieves 75.5% on MMTEB Retrieval. The 1B-BF16 model scores 72.4% on RTEB, a 27% error reduction over its predecessor llama-nemotron-embed-vl-1b-v2, and 71.0% on MMTEB Retrieval, a 28% reduction. In research agent tests powered by Nemotron 3 Ultra across ViDoRe V3, BRIGHT, and BrowseComp-Plus, the 8B model achieved the highest retrieval accuracy and lowest subsequent token cost per query. The NVFP4 variant also outperforms smaller alternatives like Qwen3-Embedding-0.6B and EmbeddingGemma-300M in serving efficiency.

Industry Impact & Enterprise Adoption

The 90% reduction in token consumption is a critical factor for Arabic-speaking teams building RAG systems on long Arabic texts, as inference costs drop significantly with the 32k context window. Multilingual retrieval enables searching across Arabic and English documents together without re-indexing. For enterprises, the integration of five technologies into a single API simplifies deployment and leverages Box's existing metadata and permission controls, making it difficult to replicate from scratch. The availability of open-weight models on Hugging Face, NIM, and vLLM lowers the barrier to entry for developers and organizations seeking to build custom retrieval-augmented generation pipelines.

Conclusion

Box's integration of NVIDIA Nemotron 3 Embed into Box Content Retrieval marks a significant step forward for enterprise AI agents, combining speed, accuracy, and cost efficiency. With the flagship 8B model topping the RTEB benchmark and the 1B variants offering competitive performance, the models provide a robust foundation for multilingual and code retrieval. As enterprises increasingly adopt AI agents, the ability to retrieve relevant information quickly and accurately while minimizing token costs will be a key differentiator. Box Enterprise Advance customers can look forward to accessing this capability early next year, while developers can start experimenting with the open models today.

Media Source: TechCrunch AI | Official Company Statement: Original Source | Fact Verification & Analysis: AI Tools Oasis