Nemotron 3 Nano 30B A3B: The Hybrid Mamba Model Behind a Family

Nemotron 3 Nano 30B A3B is worth understanding even if you never call it directly, because most of NVIDIA's 2026 small models are built on it. Nemotron 3 Nano Omni uses it as its language backbone, and Nemotron 3.5 Lightning keeps its architecture and size. Learn how the Nano behaves and you have learned most of what matters about the family.

NVIDIA put the weights on Hugging Face on 15 December 2025, with a technical report dated 23 December. It was the first model of the Nemotron 3 generation, followed by Nemotron 3 Super in March 2026.

The architecture in plain terms

The model has 52 layers of three kinds: 23 Mamba-2 layers, 23 mixture-of-experts layers and just 6 grouped-query attention layers. Each MoE layer holds 128 routed experts plus one shared expert, and a learned router activates 6 of the routed experts per token. The technical report gives 31.6B total parameters and 3.2B active per forward pass, or 3.6B counting embeddings.

The interesting part is the ratio of Mamba to attention. An attention layer keeps a key-value cache that grows with every token in the context, which is what makes long contexts expensive in memory. A Mamba-2 layer carries a fixed-size state instead. With only six attention layers, the Nano's memory cost grows far more slowly with context length than a pure transformer of similar size. Mixture of experts explained covers the other half of the design, why 3B active parameters out of 30B makes each token cheap.

NVIDIA reports up to 3.3 times the inference throughput of similarly sized open models such as GPT-OSS-20B and Qwen3-30B-A3B-Thinking-2507. That is a vendor measurement on its own stack, but the direction is what the architecture predicts.

Long context that holds up

The Hugging Face config defaults to a 256K context, and the model supports up to 1M with extra memory and, in vLLM, an environment flag to allow the longer length. OpenRouter lists the 262,144-token figure for its hosted endpoints.

What makes the long window credible is the retrieval result. On RULER-100, NVIDIA reports 92.9 at 256K, 91.3 at 512K and 86.3 at 1M, against 89.4, 84.0 and 77.5 for Qwen3-30B-A3B-Thinking-2507. The curve falls, as every model's does, but slowly.

Retrieval is not reasoning over the whole document, though. On AA-LCR, which tests reasoning across long inputs, the Nano scores 35.9 against 59.0 for the same Qwen model. It can find the needle; it is weaker at thinking about the haystack. Long context degradation explains why those two measures diverge.

Reasoning switch and thinking budget

Reasoning is on by default: the model writes a trace before its answer. Setting enable_thinking=False in the chat template turns it off, which the card says slightly reduces accuracy on harder prompts. Between the two sits a reasoning_budget setting. The model tries to close its trace at the next newline after the budget, and if no newline comes within 500 tokens it is cut off at budget plus 500.

NVIDIA recommends different sampling for different jobs: temperature 1.0 and top_p 1.0 for reasoning, and temperature 0.6 with top_p 0.95 for tool calling. Using reasoning settings for tool calls is a common cause of malformed arguments, so it is worth setting per route.

Where the benchmarks are strong and weak

NVIDIA compares the Nano with Qwen3-30B-A3B-Thinking-2507 and GPT-OSS-20B. The profile is uneven in a useful way:

• Strong — IFBench 71.5 against 51.0 and 65.0; SWE-bench under OpenHands 38.8 against 22.0 and 34.0; BFCL v4 tool calling 53.8 against 46.4 for Qwen; AIME 2025 with tools 99.2.
• Even — GPQA without tools 73.0 against 73.4 and 71.5; TauBench V2 average 49.0 against 47.7 and 48.7.
• Weak — multilingual MMLU-ProX at 59.5 against 77.6 and 69.1; Scale AI Multi Challenge 38.5 against 44.8.

The weak multilingual score matches the card's language list: English, German, Spanish, French, Italian and Japanese, with English and code as the main focus. For other languages, test carefully or pick another model.

On Artificial Analysis's Intelligence Index the Nano scores 15. That figure comes from Artificial Analysis's August 2026 launch comparison for its successor, against a field that had moved on in eight months, which leads to the real question.

Should you still use it now Lightning exists?

For new work, usually not. Nemotron 3.5 Lightning has the same size and architecture, scores 24 on the same index, improves sharply on agent benchmarks, and ships under the more permissive OpenMDW-1.1 licence rather than the NVIDIA Nemotron Open Model License that covers the Nano.

The Nano still makes sense in three cases:

1. You already fine-tuned it. A domain fine-tune on the Nano is not portable to Lightning without retraining, and a working model beats a theoretically better one.
2. You need a well-trodden local build. The card notes quantised versions for llama.cpp, Ollama and LM Studio, and tested hardware ranges from A100 and H100 to DGX Spark and Jetson Thor.
3. You want the base checkpoint. NVIDIA published both the pre-trained base and the post-trained model, which is useful for research and custom post-training.

On price there is little in it. OpenRouter lists the Nano at $0.06 per million input tokens and $0.24 per million output at the time of writing, close to Lightning's rates, so cost does not argue for staying.

Running it yourself

The BF16 weights fit on a single 80GB data-centre GPU, and the small active count keeps decode fast. vLLM, SGLang and TensorRT-LLM all serve it, using the qwen3_coder tool-call parser. Plan memory around the context you actually need rather than the 1M maximum; the Mamba layers help, but the six attention layers still keep a cache. Models that fit on one GPU puts it alongside the other single-card options.

If you are choosing a small open model today, start your evaluation with Lightning and keep the Nano as the baseline it has to beat on your own tasks. If you are reading this because a model you use is built on the Nano, its long-context and instruction-following strengths, and its multilingual weakness, are the traits to expect downstream.