Llama Nemotron
Open reasoning model family with on/off reasoning toggle
About
Llama Nemotron is NVIDIA's family of open reasoning models announced at GTC 2025, distilled and pruned from Llama 3.1/3.3 into three sizes: Nano (8B), Super (49B), and Ultra (253B). Every model ships with a reasoning toggle, so the same weights can run in fast chat mode or slow thinking mode — NVIDIA's own numbers show the Nano jumping from 36.6% to 95.4% on MATH500 when reasoning is enabled. Training combined distillation, supervised fine-tuning, and RL alignment. They are released under NVIDIA's Open Model License, which is permissive for enterprise use but not OSI-approved, so license-sensitive teams should review it. Ultra targets multi-GPU servers, so most users will reach for Nano or Super.
Specs
Context window
Pricing
Free quota
API access
Weights
Released
Benchmarks
NVIDIA-reported: Super MATH500 96.6 with reasoning on; Nano MATH500 95.4 with reasoning on (36.6 off)
Provider-reported unless a source says otherwise. Benchmarks are marketing-adjacent by nature — treat them as directional.
Use cases
Trust score — 8/10
Scored 8/10 (verified tier, metrics-v1). Evidence: NVIDIA GTC25 announcement March 18 2025 read: Nano 8B / Super 49B / Ultra 253B, reasoning on/off toggle.; Benchmarks NVIDIA-reported: Super MATH500 96.6, Nano 95.4 with reasoning on.; Training: distillation + NAS pruning + SFT + RL per arXiv 2505.00949.; License caveat: NVIDIA Open Model License, not OSI-approved — flagged.; Limitation: Ultra needs multi-GPU; most users use Nano/Super..