Nemotron 3 Nano: Open, Efficient Mixture-of-Experts Hybrid Mamba-Transformer Model for Agentic Reasoning

Nemotron 3 Nano 30B-A3B is a 52-layer model that mixes three layer types. 23 are Mamba-2 layers, a state-space design that carries a fixed-size state instead of a growing cache. 23 are mixture-of-experts layers with 128 routed experts (6 active) and 2 shared ones. Only 6 are attention layers, each with 32 query heads sharing 2 key-value heads. The model has 31.6B parameters, 3.2B active per token (3.6B with embeddings), and no positional embeddings.

It was pre-trained on 25 trillion tokens in two phases (23.5T diverse, then 1.5T high quality), with a 121B-token long-context phase mixing 512K and 4K sequences. Post-training has three parts: SFT on 18M samples, with reasoning removed from 10% to teach an on/off switch and cut short in 3% to teach budget control; reinforcement learning on all environments at once (math, code, question answering, JSON schemas, instruction following, long context, tool use); and RLHF with a 235B generative reward model. A final FP8 version keeps the attention layers and the Mamba layers feeding them in bf16.

On one H200, with 8K tokens in and 16K out, it produces 3.3 times the output tokens per second of Qwen3-30B-A3B-Thinking-2507 and 2.2 times GPT-OSS-20B. It leads both on chat (Arena-Hard-V2), instruction following (IFBench), coding (LiveCodeBench v6, SWE-Bench) and RULER up to 1M tokens. It trails Qwen3 on MMLU-Pro and on AA-LCR, and both rivals on multilingual MMLU-ProX. Weights, most of the data and the training recipe are released.

For agent and reasoning workloads, serving cost is set by how many concurrent requests fit on a GPU, and that is set by the KV cache. A reasoning model that writes 16K tokens per answer, or an agent that carries 100K tokens of history, fills memory with cache long before it runs out of compute.

Nemotron 3 Nano attacks that directly. With 6 attention layers and 2 KV heads each, the cache is about 3 KiB per token in FP8, computed from the report's Table 1; the same shape in all 52 layers would be 8.7 times larger. A million-token session needs about 3 GiB of cache. This is the mechanism behind its throughput lead on generation-heavy work, and why 1M-token context is practical on one GPU.

It is also a reference point for choosing a mid-size open model. It wins on tool use, coding and chat, and its reasoning switch and budget control were trained in rather than bolted on. It is weaker where knowledge breadth matters: MMLU-Pro, and especially multilingual MMLU-ProX (59.5 against 77.6 for Qwen3). A customer serving English agentic workloads is its sweet spot; a multilingual knowledge assistant is not.