Nemotron Cascade 2

Nemotron-Cascade-2-30B-A3B is NVIDIA's post-training research release of March 2026: a 30B-total, 3B-active MoE that runs in both instruct and thinking modes and is post-trained from Nemotron 3 Nano 30B-A3B-Base, NVIDIA's own 25T-token hybrid Mamba-2 + MoE + attention model, so it is not a new pretraining run. The checkpoint keeps the Nemotron-H shape: 52 layers in an interleaved Mamba/MoE/attention pattern, 128 routed experts with top-6 routing plus one shared expert, 32 query and 2 KV heads, a 131K vocabulary, and a 1M-token context. Weights ship under the NVIDIA Open Model License with the SFT and RL training sets released alongside.

The recipe is SFT on a curated dataset, then Cascade RL (sequential, domain-by-domain reinforcement learning, expanded from Nemotron-Cascade 1 to cover far more reasoning and agentic domains), with multi-domain on-policy distillation inserted between RL stages: the strongest intermediate checkpoint for each domain serves as that domain's teacher, which recovers benchmark regressions as later RL environments get harder. The headline result is olympiad-level reasoning at 3B active parameters: gold-medal scores on IMO 2025 (35 points), IOI 2025 (439.3), and the 2025 ICPC World Finals (10 of 12), which NVIDIA says makes it the second open-weight model after DeepSeek-V3.2-Speciale (671B) to reach all three, with 20× fewer parameters. Versus its Nemotron 3 Nano base it lifts LiveCodeBench v6 from 68.3 to 87.2, AIME 2025 from 89.1 to 92.4 (98.6 with tools), ArenaHard v2 from 67.7 to 83.5, IFBench from 71.5 to 82.9, SWE-bench Verified (OpenHands) from 38.8 to 50.2, and τ²-Bench from 49.0 to 58.9. It beats Qwen3.5-35B-A3B and the larger Nemotron 3 Super on math, code reasoning, and instruction following, but trails both on knowledge (MMLU-Pro 79.8 vs 85.3 and 83.7) and on agentic tool use (BFCL v4 52.9, Terminal-Bench 2.0 21.1). AA Intelligence Index v4.3: 12, above Nemotron 3 Nano's 9 and below Nemotron 3.5 Lightning's 16.