Model Drop

Nemotron 3 Nano

December 15, 2025 2 min read

Nemotron 3 Nano, which NVIDIA shipped on December 15, 2025, is a 30B open-weight hybrid Mamba-Transformer mixture-of-experts model with 3.5B active parameters, and the first model of the Nemotron 3 family.

Model: Nemotron 3 Nano (nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16)
Maker: NVIDIA
Ship date: December 15, 2025
Pricing: Free (open weights)
Context: 1M tokens (256K in the default config)
Where you can use it:
- Hugging Face open weights
- build.nvidia.com, NIM and hosts like OpenRouter
When to reach for it:
- Many parallel agents on one GPU
- Long codebases and document piles
- Skip it if you need image or audio input
How it compares:
- More throughput than gpt-oss-20b, per NVIDIA
- On par with Qwen3-30B-A3B-Thinking-2507
What makes it different:
- Mamba-2 layers mixed with MoE and attention
- Open weights, data, recipes and RL gyms
- Reasoning on/off plus a thinking budget
Early reactions: An early hands-on test from Ori on December 17 praised its generation speed, though it still flubbed a simple letter-counting prompt.
Read more: NVIDIA Debuts Nemotron 3 Family of Open Models, Inside NVIDIA Nemotron 3, Hugging Face model card, Nemotron 3 Nano technical report

The benchmarks

Scores are from NVIDIA's model card table against Qwen3-30B-A3B-Thinking-2507 and gpt-oss-20b; NVIDIA says RULER ran with reasoning off and AA-LCR with reasoning on, and doesn't label the rest (reasoning is on by default). NVIDIA took rival scores from official reports or Artificial Analysis and measured some itself.

NVIDIA's numbers

AIME 2025 (competition math, no tools): 89.1%. +4.1 over Qwen3-30B-A3B-Thinking-2507 (85.0%), −2.6 under gpt-oss-20b (91.7%).
LiveCodeBench (fresh coding problems, v6): 68.3%. +2.3 over Qwen3-30B-A3B-Thinking-2507 (66.0%), +7.3 over gpt-oss-20b (61.0%).
Humanity's Last Exam (no tools) (expert-level questions): 10.6%. +0.8 over Qwen3-30B-A3B-Thinking-2507 (9.8%), −0.3 under gpt-oss-20b (10.9%).
SWE-Bench (OpenHands) (real GitHub fixes): 38.8%. +16.8 over Qwen3-30B-A3B-Thinking-2507 (22.0%), +4.8 over gpt-oss-20b (34.0%).
τ²-bench Telecom (customer-support tool use): 42.2%. +15.9 over Qwen3-30B-A3B-Thinking-2507 (26.3%), −7.5 under gpt-oss-20b (49.7%).
IFBench (instruction following): 71.5%. +20.5 over Qwen3-30B-A3B-Thinking-2507 (51.0%), +6.5 over gpt-oss-20b (65.0%).
RULER-100@1M (long-context retrieval, 1M tokens): 86.3%. +8.8 over Qwen3-30B-A3B-Thinking-2507 (77.5%).
AA-LCR (long-document reasoning): 35.9%. −23.1 under Qwen3-30B-A3B-Thinking-2507 (59.0%), +1.9 over gpt-oss-20b (34.0%).

Independent numbers

Artificial Analysis Intelligence Index (current v4.3.2): 8.9. −0.9 under Qwen3-30B-A3B-Thinking-2507 (9.8), −0.1 under gpt-oss-20b (9.0).