Nemotron 3 Nano 30B A3B

Nemotron 3 Nano 30B A3B

Quick answer: NVIDIA Nemotron 3 Nano 30B A3B is an open-weight large language model built by NVIDIA on a hybrid Mamba-2 / Transformer Mixture-of-Experts architecture. It holds 30B total parameters but activates only ~3.5B per token, supports a context window up to 1M tokens, and is designed as a unified reasoning and non-reasoning model for building efficient AI agents.

What Nemotron 3 Nano 30B A3B Is

Nemotron 3 Nano 30B A3B is NVIDIA's efficiency-focused entry in the Nemotron 3 family — a model trained from scratch and released under open weights, training data, and recipes. The defining trait is its sparsity: of its 30 billion total parameters, only about 3.5 billion are active on any given token (the "A3B" tag), which is what lets a 30B-class model run on a single H100 while keeping inference cost closer to a small model.

Architecturally it departs from the standard Transformer stack. The model is a hybrid Mamba-2 / Transformer Mixture-of-Experts, with 52 layers split across 23 Mamba-2 layers, 23 MoE layers, and just 6 grouped-query attention layers. Each MoE layer carries 128 routed experts plus one shared expert, routing 6 experts per token. The Mamba-heavy design is what makes its long-context behaviour — up to a 1M-token window — practical rather than theoretical.

It is a unified reasoning model: it can emit an explicit reasoning trace before its final answer, or skip that trace via a chat-template flag when latency matters more than accuracy on hard prompts. For Benchgen's purposes, the interesting question is not its leaderboard headline but how that toggleable reasoning and tool-calling behaviour holds up across repeated runs on real, domain-specific agent tasks — exactly what FinArena is built to measure.

Specifications

Public Benchmark Scores

Scores above are reported by NVIDIA and shown for context. They are not Benchgen measurements.

How Nemotron 3 Nano 30B A3B Performs on Real Agent Tasks

The model's self-reported numbers tell a consistent story: strong reasoning (89.1 on AIME 2025, 73.0 on GPQA) and solid agentic tool use (49.0 average on τ-bench v2, 53.8 on BFCL v4) for a model this size. Those are general-capability and synthetic-environment scores, though — they describe potential, not verified behaviour on a specific domain under repeated trials.

That gap is precisely where Benchgen's FinArena measurement adds something the model card cannot. A τ-bench average says the model can call tools in a controlled retail/airline/telecom sandbox; a FinArena pass^k score says how reliably it completes a real financial agent task across k attempts, with a verified provenance trail and a failure-mode breakdown. When a FinArena run exists, this section will compare the two directly — the kind of synthesis no benchmark aggregator reproduces.

Nemotron 3 Nano 30B A3B vs Alternatives

Its most direct comparison is Qwen3 30B-class MoE models, which sit in the same active-parameter and deployment tier and which NVIDIA itself benchmarks against. The differentiators are architectural — the Mamba-heavy hybrid stack and the genuine 1M-token window — and licensing, with NVIDIA's open model license and fully released training data. For agent builders choosing a single-GPU model, the decision usually comes down to long-context efficiency (Nemotron's strength) versus ecosystem maturity (Qwen's strength).

Frequently Asked Questions

What is Nemotron 3 Nano 30B A3B? It is an open-weight large language model from NVIDIA using a hybrid Mamba-2 / Transformer Mixture-of-Experts architecture, with 30B total parameters and ~3.5B active per token, built for efficient agentic reasoning.

What is Nemotron 3 Nano's context window? Up to 1 million tokens, though the default Hugging Face configuration ships at 256k to reduce VRAM requirements.

Is Nemotron 3 Nano open source? The weights, training data, and recipes are released under the NVIDIA Nemotron Open Model License, and NVIDIA states the model is ready for commercial use.

Specs sourced from the Hugging Face model card. Third-party benchmark scores attributed inline to NVIDIA's report. Benchgen scores, where shown, are produced on Benchgen infrastructure. Last updated 2026-06-17.