Perplexity stepping in as a founding member of NVIDIA’s Nemotron Coalition is a pretty big signal about where AI is headed: open, collaborative, and deeply infrastructure‑level rather than just another shiny chatbot feature. It plugs Perplexity directly into a global effort to build frontier‑class open models that anyone can inspect, fine-tune, and deploy, instead of relying only on closed systems controlled by a handful of players.

At the heart of this move is the Nemotron Coalition itself, a new NVIDIA‑led alliance of leading AI labs created to advance “open frontier models” — think GPT-scale systems, but with open weights and transparent training practices. The coalition was announced at NVIDIA’s GTC event and aims to pool research, data, evaluation frameworks, and compute so that building these huge models becomes a shared infrastructure project instead of something every company has to reinvent alone. Members range from research‑heavy outfits like Mistral AI and Sarvam AI to applied players such as Perplexity, Cursor, and others that bring real‑world workloads and benchmarks into the loop.

The coalition’s first concrete deliverable is a new base model co-developed by Mistral AI and NVIDIA, trained on NVIDIA’s DGX Cloud and then open-sourced for the broader ecosystem to fine‑tune and adapt. NVIDIA has been explicit that this base model will underpin its upcoming Nemotron 4 family, effectively turning the coalition’s work into the foundation for future high‑end NVIDIA models as well. Coalition members contribute at different layers: some bring sovereign‑language and regional expertise, others provide evaluation datasets, and others inject specialized domain knowledge from production systems.

Perplexity’s role is very much on that “real usage” side of the spectrum: it already runs a complex retrieval‑augmented search product at scale and has a habit of stitching together different open models for each stage of answering a query. Under the hood, Perplexity post-trains different open models for query parsing, retrieval, reranking, and drafting responses, which lets it tune latency, cost, and relevance for each step instead of throwing one monolithic model at everything. That experience—knowing where models fail, how they behave under heavy search traffic, and which fine‑tuning knobs actually matter—is exactly the kind of domain expertise the coalition wants to bake into its shared base models.

Open models are the philosophical core of this whole effort, and that matters more than the branding. Pre-training a frontier-scale model is the most expensive, resource-intensive part of the pipeline; once that’s open, thousands of smaller teams can afford to specialize and fine‑tune instead of trying to raise billions to compete from scratch. NVIDIA’s own positioning here is blunt: it calls open models “the lifeblood of innovation” because they invite students, startups, and enterprises worldwide to participate in the AI stack rather than just consume it.

Nemotron 3 Super is a good example of the kind of open foundation Perplexity is leaning into. This is a 120-billion-parameter hybrid MoE model, but only 12 billion parameters are active at inference time thanks to a LatentMoE architecture, which makes it far more efficient than a naive 120B-dense system. Nemotron 3 Super is optimized for agentic workloads rather than simple chat: long-context reasoning, tool calling, planning, code and IT automation, and other multi-step tasks where multiple tools and data sources come into play.

Perplexity has already wired Nemotron 3 Super into its own stack in three ways: in the model selector inside the search experience, via its Agent API, and as part of Perplexity Computer, where the model is integrated directly into the search pipeline. That means a lot of real queries—research tasks, coding questions, complex multi‑step prompts—will effectively act as stress tests and real‑world feedback loops for Nemotron 3 Super and future Nemotron-line models. For the coalition, this kind of deployment is gold: it surfaces edge cases, helps refine evaluation benchmarks, and proves whether these open models can actually hold up against proprietary systems in day-to-day use.

On the NVIDIA side, Nemotron is a broader strategy to make its hardware and software stack the default home for large-scale open models. Nemotron models are designed to run especially efficiently on NVIDIA’s latest platforms—Blackwell GPUs, NVFP4 precision, and so on—which means better throughput and lower cost for anyone building on top of them. In return, NVIDIA publishes model weights, recipes, and tooling that make it easier for developers to stand up their own fine‑tuned versions, either in the cloud or on-prem.

For Mistral AI, a fellow founding member, Nemotron is a natural extension of its own open-first philosophy. Mistral is contributing cutting-edge architectures, multimodal capabilities, and large‑scale training know-how, while using NVIDIA’s compute and tooling to push these open models to frontier scale. Combined with Perplexity’s search and agent workloads, this starts to look like a full ecosystem loop: Mistral and NVIDIA push the state of the art, Perplexity and others pressure‑test it in production, and the resulting improvements flow back into the open model base for everyone.

What this all adds up to is an attempt to change the balance of power in AI infrastructure without pretending that closed models will disappear overnight. By banding together under the Nemotron Coalition, labs like Perplexity get access to serious compute and a shared base model that would be painful to build alone—while still retaining the ability to keep their own post-training magic proprietary if they want. For developers and enterprises watching from the outside, the promise is simple but ambitious: frontier‑grade models that are open enough to inspect and customize, battle‑tested on real workloads like Perplexity’s, and backed by some of the strongest infrastructure in the industry.