NVIDIA Nemotron 3 Nano 30B: 3B Active Parameters and Released RL Environments

What Nemotron 3 Nano 30B A3B is

NVIDIA Nemotron 3 Nano is a 30 billion parameter mixture-of-experts model with 3.5 billion active per token, per NVIDIA's card, published in BF16 under the NVIDIA Nemotron Open Model License with a 256K token context window. At roughly $0.06 per million input tokens and $0.24 output through third-party providers it is among the cheapest models here.

NVIDIA publish the environments, not just the weights

NVIDIA state that the model underwent multi-environment reinforcement learning using synchronous GRPO — Group Relative Policy Optimization — across mathematics, code, science, instruction following, multi-step tool use, multi-turn conversations and structured output environments, with conversational quality further refined through RLHF using a generative reward model they also publish. Every dataset is disclosed in the card's own training and evaluation section. And the reinforcement-learning environments themselves are released as NeMo Gym.

NVIDIA’s comparison table

From NVIDIA's own comparison table: on SWE-Bench under OpenHands it records 38.8, the best of the three models compared, against 22.0 and 34.0. On BFCL v4, a function-calling benchmark, it takes 53.8 against 46.4. On the TauBench V2 average it leads at 49.0, though the per-domain picture is mixed — it wins telecom's comparison at 42.2 only in one column's reading and trails on airline at 48.0 against 58.0 and retail at 56.9 against 58.8. On WMT24++ translation it leads at 86.2. On MMLU-ProX averaged over languages it records 59.5 against 77.6. Terminal Bench on its hard subset is low for everything compared, at 8.5 against 5.0 and 6.0.

LMArena

On LMArena's text leaderboard NVIDIA Nemotron 3 Nano 30B A3b Bf16 has 1314 Elo (rank 274); by category, 1361 Elo in Coding (rank 259), 1348 Elo in Math (rank 220), 1332 Elo in Expert (rank 237) and 1242 Elo in Creative Writing (rank 304).

Licence

It is released under the NVIDIA Nemotron Open Model License, which is not Apache 2.0. The NeMo Gym release contains the reinforcement-learning environments used in training.

Architecture and training

NVIDIA describe a hybrid Mamba-2 and Transformer mixture-of-experts design with 52 layers: 23 Mamba-2 layers, 23 MoE layers and 6 grouped-query attention layers. Each MoE layer has 128 routed experts plus one shared expert, with 6 activated per token. It was trained from scratch on 25 trillion tokens with a batch size of 3,072 and a warmup-stable-decay learning-rate schedule; the pre-training data cutoff is 25 June 2025 and the post-training cutoff 28 November 2025. It was released on 15 December 2025. Reasoning can be switched off with a chat-template flag, which NVIDIA say slightly reduces accuracy on harder prompts. Supported languages are English, German, Spanish, French, Italian and Japanese, and NVIDIA report 86.3 on RULER at 1 million tokens.

NVIDIA’s published results

NVIDIA's table compares the model with Qwen3-30B-A3B-Thinking-2507 and GPT-OSS-20B. For Nemotron 3 Nano it reports MMLU-Pro 78.3, GPQA 73.0 without tools and 75.0 with tools, SWE-Bench (OpenHands) 38.8, TauBench V2 average 49.0, BFCL v4 53.8, Arena-Hard-V2 average 67.7 and MiniF2F pass@1 50.0. NVIDIA list Qwen3-30B-A3B-Thinking-2507 at 80.9 on MMLU-Pro and 22.0 on SWE-Bench, and GPT-OSS-20B at 91.7 on AIME25 without tools. Evaluations were run with the NeMo Evaluator SDK and NeMo Skills, and the configurations are published in the repository. Training ran in three stages: pre-training with Megatron-LM, supervised fine-tuning on synthetic code, maths, science and tool-calling data, and multi-environment reinforcement learning with synchronous GRPO in NeMo RL and NeMo Gym.

Official Resources