Nemotron 3 Nano

NVIDIA · 2025-12-23 · analysed 2026-08-10 · skill v0.3.5

sparse-MoE SSM text-only open-weights-open-data large reasoning-specialised

Design overview

Nemotron 3 Nano 30B-A3B is an open, efficient sparse MoE hybrid Mamba-Transformer: 52 layers, 31.6B total parameters with only 3.2B active per forward pass (3.6B including embeddings). The backbone follows the Nemotron-H / Nemotron 2 Nano lineage — a Jamba-style interleaving of Mamba-2 state-space layers and Grouped-Query-Attention layers, 6 of 52 layers being GQA (32 query heads, 2 KV heads, head dimension 128) and the remaining 46 being Mamba-2 (state dimension 128, 8 groups, 64 heads) — with every standard FFN replaced by a granular MoE: 128 routable experts of dimension 1856, 6 activated per token plus 2 shared experts, a learnt MLP sigmoid router, squared ReLU activation, and DeepSeek's aux-loss-free bias-based load balancing combined with a small auxiliary loss. The model is fully NoPE: no positional embeddings, no dropout, no linear-layer bias, RMSNorm, untied embeddings. Context reaches 1M tokens through a final LC-Phase of continuous pretraining on a 512k/4k sequence mixture, with no positional-encoding extension required. Post-training runs SFT over chat, agentic, and reasoning traces with reasoning on/off and token-budget control, a unified multi-environment RLVR stage (synchronous GRPO with masked importance sampling and frozen router weights), and RLHF with a GRPO-trained generative reward model plus Group Relative Length Control that cut verbosity 30% without accuracy loss. Deployment ships BF16 and a selective FP8 PTQ checkpoint that keeps the 6 attention layers and their feeding Mamba layers in BF16, recovering ~99% median accuracy.

What the report claims is novel

Atlas assessment: Architecturally an incremental combination of established mechanisms: the Mamba-2 + GQA hybrid follows the Jamba / Nemotron-H / Nemotron 2 Nano lineage, the granular MoE with shared experts follows DeepSeekMoE, and the aux-loss-free bias-based load balancing follows DeepSeek-V3. The genuinely distinctive contributions are in post-training and efficiency: unified multi-environment RLVR (all environments simultaneously), RLHF with a generative reward model plus Group Relative Length Control (a new, simple reward shaping that cut verbosity 30% with no accuracy loss), and a selective FP8 PTQ scheme (6 attention + 6 feeding Mamba layers kept BF16). Squared ReLU MoE at 6/128 routing is a notable but minor variant. Nothing here is a new mixing mechanism; the report's headline is the accuracy/throughput trade-off for an open 30B-class agentic model.

Relation to the atlas

Lineage in

Influence out

Notable omissions

Closest relatives in the atlas

Spec sheet

Identity

FamilyNemotron 3 Nano[1]
OrganisationNVIDIA[2]
Release date2025-12-23[3]
Report URLhttps://research.nvidia.com/labs/nemotron/files/NVIDIA-Nemotron-3-Nano-Technical-Report.pdf[4]
Licensenot disclosed
Open weightsyes[5]

Scale

Total parameters31.6B[6]
Active parameters / token3.2B[7]
Layers52[8]
Hidden dimension2688[9]
FFN inner dimension1856[10]
FFN ratio0.69
Query heads32[11]
KV heads2[12]
Head dimension128[13]
Vocabulary sizenot disclosed
Embedding tyingno[14]

Core block

Block typesparse-MoE
Attention variantGQA[15]
Attention layer pattern6 of 52 layers are GQA self-attention layers, the remaining 46 are Mamba-2; Figure 2 depicts a periodic interleaving of Mamba-2+MoE and Mamba-2+Attention+MoE blocks (repeat counts x5/x3/x1/x4 per the figure) — the extracted figure text does not fully resolve the exact interleaving, but §4.2 states 6 self-attention layers and that the 6 Mamba layers feeding them are kept in BF16[16]
Depth mixingsequential-residual
Activationsquared ReLU (on the MoE layers)[17]
Expert count128[18]
Experts per token6[19]
Shared experts2[20]
Routingstandard learnt MLP router with sigmoid gating; router weights frozen during RLVR to stabilize training[21]
Load balancingDeepSeek aux-loss-free load balancing (expert-bias based) with update rate 10−3, used in conjunction with the standard load balancing loss with coefficient 10−4; sequence-level MoE load balancing regularizer (coefficient 10−4) during SFT[22]
Expert granularitygranular MoE (DeepSeekMoE-style fine-grained experts) with shared experts: 128 routable experts of dimension 1856, 6 active per token, plus 2 shared experts[23]
MethodNoPE[27]
RoPE base frequencyn/a[28]
Partial RoPEn/a[29]
TypeRMSNorm[30]
Placementnot disclosed
QK-normno
Attention sinksno
Softcappingno

Context

Trained context length524,288[31]
Deployed context length1M[32]
Extension methodLC-Phase continuous pretraining (CPT): 512k and 4k sequence mixture, 121B tokens, with 20% long-context document QA and 1% synthetic retrieval-focused data (up to 256k); no positional-encoding modification needed (NoPE)[24]

Tokenizer

Algorithmnot disclosed
Notesnot disclosed

Training

Training tokens25T[33]
Data composition15 data categories: web crawl subdivided into five quality groups (crawl-medium, crawl-medium-high, syn-crawl-medium-high, crawl-high, syn-crawl-high) plus math, Wikipedia, code, nemotron-cc-code, academic text, Crawl++ (OpenWebText, BigScience, Reddit), multilingual (19 languages), and synthetic SFT-style datasets (general-sft, stem-sft, code-sft); new released datasets: Nemotron-CC-v2.1 (~2.5T new English tokens: 3 fresh Common Crawl snapshots, 5-prompt rephrasing of Medium-High data from 110 snapshots, translation to English from 9 languages), Nemotron-CC-Code-v1 (427.92B-token code corpus via Lynx + Phi-4 cleaning), Nemotron-Pretraining-Code-v2 (GitHub refresh + Qwen3-32B synthetic dialogues, SGCR/SCOR rephrasing, Python-to-C++ transpilation), Nemotron-Pretraining-Specialized-v1 (synthetic Wikipedia, math textbooks, scientific coding, InfiniByte cross-domain code, RQA 4.3M demonstrations / ~31.7B tokens, DQA ~8B tokens)[34]
Curriculum / stagingtwo-phase curriculum: Phase 1 promotes diversity, Phase 2 primarily high-quality datasets (e.g., Wikipedia), switch at the 94% point of training; LC-Phase CPT at the end (512k/4k mixture); RLVR uses a difficulty curriculum with Gaussian pass-rate sampling that shifts from easy to hard tasks across training steps[35]
OptimizerAdamW with weight decay 0.1, β1 = 0.9, β2 = 0.95[36]
LR scheduleWarmup-Stable-Decay: warmed up over 8.4B tokens to a maximum of 10−3, maintained for 80% of training (20T tokens), decayed to a minimum of 10−5 over the final 5T tokens; LC-Phase used a constant LR of 10−5[37]
Batch schedulepretraining: sequence length 8192, batch size 3072 (~25M tokens per batch); LC-Phase: global batch size 48; SFT: batch size 64 with sequence packing to 256K; RLVR: batch size 2048 (128 prompts/step, 16 generations/prompt)[38]
PrecisionBF16[39]
ParallelismLC-Phase: 8-way context parallelism, 8-way tensor parallelism, 8-way expert parallelism, 4-way pipeline parallelism; main pretraining parallelism not disclosed[24]
HardwareH100 GPUs (disclosed only for the LC-Phase; main pretraining hardware not stated)[40]
Disclosed computenot disclosed

Post-training

SFTyes[41]
Preference optimisationRLHF[42]
Reasoning trainingmulti-environment RLVR from verifiable rewards: all environments trained simultaneously (competition math DAPO/SkyWorks, competition coding, STEM MCQ QA, structured outputs/JSON schema, instruction following, long context, agentic tool use incl. Workplace Assistant and multi-turn conversational agent); synchronous GRPO with masked importance sampling, 128 prompts/step, 16 generations/prompt, batch 2048, max generation length 49K, overlong filtering, difficulty curriculum; two RLVR stages (after SFT and after RLHF); reasoning on/off and token-budget control via the chat template[43]
Distillationnone

Modality

Typetext-only
Attachmentn/a[44]

Inference efficiency

KV-cache designGQA with 2 KV heads (KV cache ∝ 2 heads); KV cache quantized to FP8 via PTQ — FP8 KV cache significantly improves throughput by enabling larger batch sizes[45]
Quantisation shippedFP8 post-training quantization (PTQ) via ModelOpt and Megatron-LM with selective BF16 retention: the 6 self-attention layers, the 6 Mamba layers feeding them, and all Mamba Conv1D layers kept in BF16; weights, activations, and KV cache FP8; ~99% median accuracy recovery[46]
Speculative decodingnone disclosed
Serving optimisationsinference throughput measured with vLLM and TRT-LLM (best of the two per model) on a single H200 GPU with FP8 weights and activations; no dedicated serving-optimization stack described[47]

Evaluation

Benchmarks (report's own numbers only)

BenchmarkValueRef
MMLU-Pro78.3Table 3
AIME25 (no tools)89.06Table 3
AIME25 (with tools)99.17Table 3
GPQA (no tools)73.04Table 3
GPQA (with tools)75.0Table 3
LiveCodeBench (v6 2024-08↔2025-05)68.25Table 3
SciCode (subtask)33.28Table 3
HLE (no tools)10.57Table 3
HLE (with tools)15.48Table 3
MiniF2F pass@150.03Table 3
MiniF2F pass@3279.92Table 3
Terminal Bench (hard subset)8.51Table 3
SWE-Bench (OpenHands)38.76Table 3
TauBench V2 Airline48.0Table 3
TauBench V2 Retail56.91Table 3
TauBench V2 Telecom42.21Table 3
TauBench V2 Average49.04Table 3
BFCL v453.76Table 3
IFBench (prompt)71.51Table 3
Scale AI Multi Challenge38.45Table 3
Arena-Hard-V2 (Average)67.65Table 3
AA-LCR35.85Table 3
RULER-100 @ 256k92.92Table 3
RULER-100 @ 512K91.25Table 3
RULER-100 @ 1M86.34Table 3
MMLU-ProX (avg over langs)59.5Table 3
WMT24++ (en→xx)86.2Table 3

Provenance references

  1. stated · Title
  2. stated · Title page byline
  3. stated · Title page date line (2025-12-23)
  4. stated · pipeline
  5. stated · Abstract: 'We release both our pretrained Nemotron 3 Nano 30B-A3B Base and post-trained Nemotron 3 Nano 30B-A3B checkpoints on Hugging Face'
  6. stated · §2.1 ('31.6B total parameters'); Abstract
  7. stated · §2.1 ('3.2B are active (3.6B including embeddings) per forward pass'); Abstract ('activating less than half of the parameters per forward pass')
  8. stated · Table 1 (Num Layers)
  9. stated · Table 1 (Model Dimension)
  10. stated · Table 1 (Expert Dimension)
  11. stated · Table 1 (Q-heads)
  12. stated · Table 1 (KV-heads); §1 (GQA)
  13. stated · Table 1 (Head Dimension)
  14. stated · §2.1 ('un-tie embedding and projection weights')
  15. stated · §1 ('combination of Mamba-2 (Dao & Gu, 2024) and Grouped-Query-Attention (GQA) (Ainslie et al., 2023)')
  16. stated · §4.2 ('self-attention layers (6 out of 52 layers for Nemotron 3 Nano)'); Figure 2
  17. stated · §2.1 ('For the MoE layers, we use squared ReLU activation')
  18. stated · Table 1 (Total Routable Experts)
  19. stated · Table 1 (Number of Activated Experts); §1 ('activates 6 out of 128 experts')
  20. stated · Table 1 (Number of Shared Experts)
  21. stated · §2.1; §3.2.5
  22. stated · §2.4; §3.1.6; §3.2.5
  23. stated · §2.1 ('granular MoE architecture along with shared experts (Dai et al., 2024)'); Table 1
  24. stated · §2.5
  25. n/a · §2.5 (no rescaling/interpolation factor; no RoPE)
  26. stated · §1, §5 ('supports context lengths up to 1M tokens'); RULER-100 @ 1M evaluated (Table 3)
  27. stated · §2.1 ('We do not use any positional embeddings')
  28. n/a · §2.1 (no positional embeddings of any kind; no RoPE anywhere)
  29. n/a · §2.1 (no RoPE)
  30. stated · §2.1 ('We use RMSNorm for normalization')
  31. stated · §2.5 (LC-Phase CPT on batches of 524,288 (512k) token sequences, mixed with 4k sequences; base pretraining at 8192)
  32. stated · §1 ('supports context lengths up to 1M tokens'); RULER-100 @ 1M (Table 3)
  33. stated · Abstract ('pretrained on 25 trillion text tokens'); §2.4; note: 23.5T phase 1 + 1.5T phase 2 + 121B LC-Phase (§1, §2.5)
  34. stated · §2.2, §2.3
  35. stated · §2.3, §2.5, §3.2.2
  36. stated · §2.4
  37. stated · §2.4, §2.5
  38. stated · §2.4, §2.5, §3.1.6, §3.2.5
  39. stated · §4 ('After post-training the model in BF16, we applied Post-Training Quantization (PTQ) using ModelOpt and Megatron-LM to quantize the model to FP8')
  40. stated · §2.5 ('to train on H100 GPUs'); throughput measured on a single H200 GPU (§1)
  41. stated · §3.1 (SFT over 18M samples; 13000 steps, batch size 64, sequence packing to 256K, LR 5·10−5 with 800 warmup steps; chat/agentic/reasoning traces; reasoning on/off control by stripping traces from 10% of samples, budget control by truncating 3%)
  42. stated · §3.3 (RLHF with a generative reward model (GenRM) built by GRPO-training Qwen3-235B-A22B-Thinking-2507; Group Relative Length Control with length-normalized reward adjustment and quality-gated conciseness bonus; verbosity reduced 30% without accuracy loss)
  43. stated · §3.2
  44. n/a · modality.type (text-only)
  45. stated · §4.2 ('The model weights, activations, and KV cache are quantized to FP8'), §4.3
  46. stated · §4
  47. stated · §1, Figure 1 caption