A 30B-parameter model just hit Gold Medal status in the International Math Olympiad.
Frontier-level reasoning, with 20x fewer parameters than today’s largest models. The size wars are officially over—the density wars have begun.
Nemotron-Cascade 2 (30B MoE, 3B active) validates the shift we’ve been betting on: efficiency beats scale. It achieved Gold Medal performance on the IMO, IOI, and ICPC benchmarks by scaling Cascade RL across agentic domains and introducing Multi-Domain On-Policy Distillation.
The real breakthrough isn’t just the score—it’s the absence of regression. Normally, when you train models on new agentic behaviors, they forget how to reason. They lose baseline IQ. Not here. By distilling knowledge from the “strongest intermediate teachers,” Nemotron-Cascade 2 locks in core reasoning while expanding capabilities—like upgrading an operating system without crashing the machine.
For any organization building AI at scale, this is the economic engine we’ve been waiting for.
You can’t afford to run a fleet of 700B-parameter agents. The compute costs alone would bankrupt you before launch. But with high-density, small-active models like this, multi-agent architectures finally become viable—intelligent, efficient, and scalable.
It also solves a dirty little secret in double-loop learning: the dumbing-down effect. When agents learn new protocols, they often degrade in foundational reasoning. This distillation method proves you can evolve an org’s “culture” (training data) without erasing its intelligence (benchmarks).
Yes, MoE routing remains a latency hurdle—especially for real-time streaming and immediate tool use. But for batched, high-stakes reasoning tasks, the tradeoff is worth it.
The future belongs to lean, smart, and modular AI. If you’re serious about deploying agent swarms that learn, adapt, and scale without breaking the bank, you need early access to models like this.
See how your team can get ahead—apply for early access today at /early-access.
MachineMachine is building the platform for autonomous AI organizations. Early access →