June 2025 Summaries

Perplexity plans to integrate European sovereign AI models into its answer engine later this year through a partnership with NVIDIA, aiming to combine global availability with language-specific, culturally informed responses. Developed by institutions across countries including France, Germany, Italy, Poland, Spain, and Sweden, the models are intended to better understand European languages, local knowledge, and cultural contexts than broadly trained global alternatives. NVIDIA will optimize the models using Nemotron post-training methods such as neural architecture search, reinforcement learning, and synthetic data, while NVIDIA NIM microservices will support deployment across cloud, data center, and on-premises environments. Hosted on European infrastructure through NVIDIA Cloud Partners, the initiative is designed to keep data in Europe, support the region’s digital sovereignty goals, and provide enterprises, publishers, organizations, and worldwide users with answers informed by localized European expertise.

Speculative decoding accelerates large language model inference by having a smaller draft model propose several tokens that a larger target model verifies, allowing multiple accepted tokens plus an additional target-generated token to be emitted in one step. Perplexity applies this approach to reduce inter-token latency for Sonar models through several methods, including conventional target-draft pairing, EAGLE-style feature-aware prediction, and multi-token prediction (MTP) heads that use target hidden states as well as tokens. Target-draft decoding can reuse a fine-tuned small LLM but adds KV-cache memory use and prefill overhead, while EAGLE can reduce memory needs with compact custom draft layers but requires specialized training and tree attention masks that may significantly slow verification. Perplexity therefore emphasizes linear and single-token MTP variants, developing infrastructure to train prediction heads efficiently and finding that restoring RMS normalization layers improved convergence and accuracy for large models. Its inference implementation tightly couples draft and target models, shares scheduling and KV-cache allocation, and uses overlapping CPU and GPU work to limit synchronization costs, with different scheduling strategies tailored to full draft models and faster MTP layers.

Perplexity has introduced an SEC/EDGAR integration intended to make public-company filing data more accessible to investors, business owners, students, advisors, bankers, and legal professionals. The feature uses Perplexity’s answer engine to interpret financial statements, strategic plans, and material risk disclosures from SEC filings, while providing direct source citations for verification and further research. Positioned as an alternative to costly, complex traditional financial-data platforms, the integration can be used with Perplexity Search, Research, and Labs to combine filings with market news, industry analysis, and peer comparisons. Enterprise Pro customers can also search SEC data alongside FactSet M&A and transcript information, Crunchbase firmographic data, and internal company files.