NVIDIA Nemotron-Cascade 2 debuts open 30B MoE

Hot take: this looks like NVIDIA pushing sparse MoE as the practical path to "big-model" capability without big-model cost. The real question is whether it delivers outside benchmark slides, but the packaging is strong enough that developers will actually try it.

• –The model card says it is an open 30B MoE with 3B active parameters, released under NVIDIA's Open Model License on March 19, 2026.
• –NVIDIA's claims lean hard on math, code, and agentic evals, including IMO, IOI, ICPC, LiveCodeBench, and SWE Verified performance.
• –Early Reddit reaction looks split between "interesting and usable" and "sparse MoE counts can be misleading," which is the right kind of skepticism for this release.
• –One commenter says it runs on 6GB VRAM plus system RAM and is useful for background docs and small coding tasks, which suggests real local-deployment appeal.
• –The open chat template and tool-use prompting make it easier to slot into agent workflows than a raw benchmark drop.