Nemotron Cascade 2 tops math, code

// 68d agoMODEL RELEASE

Nemotron Cascade 2 tops math, code

NVIDIA’s open 30B MoE reasoning model uses just 3B active parameters and is post-trained from Nemotron-3-Nano-30B-A3B-Base. NVIDIA says it reaches gold-medal-level results on IMO, IOI, and ICPC-style tasks, while supporting both thinking and instruct modes.

// ANALYSIS

NVIDIA is making a strong case that post-training, not brute scale, is what buys frontier-ish performance here.

–The benchmark table suggests this 30B model can hang with or beat NVIDIA’s 120B-class sibling on several math and code tasks.
–Cascade RL plus multi-domain on-policy distillation looks like a reusable recipe for squeezing more capability out of open MoE models.
–Thinking and instruct modes make it more flexible for agent workflows, not just benchmark demos.
–Open weights plus released data/checkpoints make it unusually attractive for researchers and local model tinkerers.

// TAGS

nemotron-cascade-2-30b-a3bllmreasoningbenchmarkresearchopen-weightsopen-sourceai-coding

DISCOVERED

68d ago

2026-03-20

PUBLISHED

68d ago

2026-03-20

RELEVANCE

10/ 10

AUTHOR

Middle_Bullfrog_6173

// KEEP READING

More AI developer news from the feed

EXPLORE FULL FEED

UPDATE3h ago

Cursor adds dedicated subagents for skills

Cursor now allows developers to execute tool-heavy or research-intensive agent skills within dedicated subagents. This architectural shift isolates noisy background tasks, keeping the main chat context clean and focused.

UPDATE3h ago

YouTube moves AI labels to video player

YouTube is moving its AI content disclosures from video descriptions to more prominent placements beneath the player and on Shorts overlays. Starting in May, the platform will use internal signals to automatically label photorealistic AI content that creators fail to disclose.

OPEN SOURCE7h ago

Taste Skill kills AI "frontend slop"

Taste-Skill is an open-source framework that provides portable "agent skills" to enforce high-end design principles in AI-generated code. By injecting specific design directives and "anti-slop" rules, it enables LLMs to produce editorial-grade UIs that bypass generic, boilerplate-heavy AI templates.