Perplexity adds NVIDIA's Nemotron 3 Super across its lineup

Anyone who has built a multi-agent system knows the pain: agents passing conversation history back and forth quickly leads to context explosion, and burning massive frontier models on routine sub-tasks racks up a brutal thinking tax.

NVIDIA just rolled out Nemotron 3 Super, an open-weights model packing 120 billion total parameters with only 12 billion active per token. Perplexity immediately jumped on it, rolling out the model across its main app, Agent API, and Perplexity Computer workspace.

Mamba meets Transformer in a hybrid MoE

To keep performance high without spiking inference latency, NVIDIA stitched together three distinct techniques into a single backbone. It interleaves Mamba-2 layers for linear-time sequence handling with Transformer attention layers for precise associative recall, all wrapped inside a mixture-of-experts setup.

The real trick here is what NVIDIA calls Latent MoE. Instead of sending raw, high-dimension token embeddings directly to expert networks, the model projects them into a compressed latent space first. This compression lets the system activate four times as many expert specialists at runtime for roughly the same computational cost.

Combined with multi-token prediction and native NVFP4 pretraining tuned for the Blackwell architecture, Nemotron 3 Super delivers more than five times the throughput of the previous Nemotron Super. On NVIDIA B200 hardware, inference runs up to four times faster than FP8 on H100s. To prevent multi-turn workflows from drifting off-target, it features a native 1-million-token context window, earning an 85.6% score on PinchBench.

Perplexity puts it straight to work

Perplexity wasted no time integrating the model into its production stack. End users can select it for web search, developers can target it via the Agent API, and it now sits as one of 20 orchestrated engines inside Perplexity Computer.

Outside of Perplexity, developer tools like CodeRabbit, Greptile, and Factory are wiring Nemotron 3 Super into their automated coding and debugging pipelines to cut token bills. Heavyweight enterprise software players, including Palantir, Siemens, and Cadence, are also adopting the architecture for tasks spanning semiconductor layout, cybersecurity triage, and factory automation.

NVIDIA also took the open-source route, publishing model weights alongside over 10 trillion tokens of training data and evaluation recipes.

Why it matters

For engineers building autonomous workflows, Nemotron 3 Super looks like one of the most practical open-source backbones released so far. Multi-agent designs routinely choke when passing dense logs and execution trees between sub-agents. A model that activates just 12 billion parameters while holding a million tokens in active memory lets developers ingest entire codebases without expensive chunking or latency spikes.

For the wider industry, this release shows NVIDIA tightening the grip between its silicon and frontier model design. By co-designing an unorthodox Mamba-Transformer hybrid around the Blackwell NVFP4 format, NVIDIA is giving developers an aggressive price-performance ratio that actively incentivizes staying on its newest data center chips.

We think targeting agent throughput rather than chasing generic chat benchmarks is a smart, pragmatic play. The catch, however, is hardware access: while the open weights run anywhere, the headline fourfold speed gains heavily rely on Blackwell, leaving teams on older Hopper clusters with a narrower slice of that promised efficiency.