Nemotron 3 Ultra (Nemotron 3 family)

NVIDIA · 2026-06-09 · analysed 2026-08-10 · skill v0.3.5

sparse-MoE hybrid text-only open-weights-open-data frontier frontier-generalist

Design overview

Nemotron 3 Ultra is a 108-layer, 550B-parameter sparse MoE (55B active per token) built as a scaled-up Nemotron 3 Super: a hybrid Mamba-Attention backbone in which Mamba-2 state-space layers dominate and sparse global-attention anchors (GQA, 64 query heads / 2 KV heads, head dim 128) provide periodic exact-attention interaction, with every layer carrying a LatentMoE FFN (512 experts per layer, top-22 activated, experts operating in a 2,048-wide latent space instead of the full 8,192 width, plus a shared expert with 10,240 intermediate size). Pre-training runs 20T tokens in NVFP4 — E2M1 weights with 2D block quantization and stochastic rounding, with the final 16 layers and all attention/latent/embedding/MTP projections held in higher precision — under a Warmup-Stable-Decay schedule (200B warmup to 2.5e-4, minus-sqrt decay over the final 5T), split into a 15T diversity phase and a 5T quality phase. Two training divergences are dissected in detail: a BF16 gradient-accumulation regression for the output layer (fixed by returning to FP32 reduction) and an undetermined divergence at ~16T (mitigated by early annealing to a 20T horizon). A 33B-token LC-Phase continuous-pretraining stage (92% of iterations at 1,048,576 tokens, 8% at 4K) delivers the 1M context. Post-training is agent-focused: two-stage SFT, unified RLVR, then two iterations of asynchronous Multi-teacher On-Policy Distillation from more than ten specialized teachers (SWE, office, search, terminal-use, chat, safety, STEM, coding…), finished with MTP Boosting that aligns the shared-weight two-block draft head to backbone logits for speculative decoding. Reasoning-effort control (off/regular/medium) tunes the accuracy-compute trade-off at inference.

What the report claims is novel

Atlas assessment: The block is a scaled-up Nemotron 3 Super: hybrid Mamba-2 + sparse GQA attention with LatentMoE FFNs, so within the atlas its closest relatives are ai21-jamba (hybrid SSM-attention, dense) and deepseek-v3 (frontier MoE + MTP) rather than any all-attention MoE. Biggest design differences vs deepseek-v3: the Mamba-2 majority backbone (attention only as sparse anchors, 2 KV heads) instead of all-MLA layers, NVFP4 pre-training instead of BF16/FP8, and the two-iteration MOPD pipeline instead of single-model RLVR. MTP-as-native-drafter and NVFP4 5.03-BPE mixed quantization are shared with the Nemotron 3 family rather than novel vs the atlas; MOPD (on-policy distillation of multiple teachers) mirrors the kimi-k3 and nemotron-cascade-2 direction and is not new to the atlas. The 550B/55B scale with a 512-expert LatentMoE (22 active) and 1M context is the largest hybrid-SSM record in the atlas. Claims of ~6× throughput are measured in-house on GB200 (TRT-LLM vs vLLM baselines) and are not independently verifiable.

Relation to the atlas

Lineage in

Influence out

No descendants recorded yet.

Notable omissions

Closest relatives in the atlas

Spec sheet

Identity

FamilyNemotron 3 Ultra (Nemotron 3 family)[1]
OrganisationNVIDIA[2]
Release date2026-06-09[3]
Report URLhttps://research.nvidia.com/labs/nemotron/files/NVIDIA-Nemotron-3-Ultra-Technical-Report.pdf[4]
Licensenot disclosed
Open weightsyes[5]

Scale

Total parameters550B[6]
Active parameters / token55B[7]
Layers108[8]
Hidden dimension8192[9]
FFN inner dimension5120[10]
FFN ratio0.625
Query heads64[11]
KV heads2[12]
Head dimension128[13]
Vocabulary sizenot disclosed
Embedding tyingnot disclosed

Core block

Block typesparse-MoE
Attention varianthybrid[14]
Attention layer patternsparse global Attention layers anchored among Mamba-2 layers: repeating groups of Mamba-2 + LatentMoE blocks with periodic Attention + LatentMoE anchors (Figure 2); exact per-layer counts are not stated in the prose[15]
Depth mixingsequential-residual
Activationnot disclosed
Expert count512[16]
Experts per token22[17]
Shared expertsnot disclosed
Routingnot disclosed
Load balancingnot disclosed
Expert granularityfine-grained LatentMoE: routed experts operate in a compact latent space (MoE Latent Size 2048) rather than at full hidden width; 512 experts per layer with top-22 active
Methodnot disclosed
RoPE base frequencynot disclosed
Partial RoPEnot disclosed
Typenot disclosed
Placementnot disclosed
QK-normnot disclosed
Attention sinksno
Softcappingno

Context

Trained context length1,048,576[19]
Deployed context length1M[20]
Extension methodLC-Phase continuous pretraining (CPT): 1M context for 92% of iterations and 4K for 8% (math/code SFT-style data only), constant LR 2.5e-6, 33B tokens; long-context data 46% + Phase-2 data 54% in the blend; no RULER-style data[21]

Tokenizer

Algorithmnot disclosed
Notesnot disclosed

Training

Training tokens20T[22]
Data compositionTwo-phase mixture (adaptation of the Nemotron 3 Super/Nano recipe): phase 1 (~15T tokens, ~75%) biases diversity, phase 2 (~5T) biases quality; 19 high-level categories, largest being quality-filtered and synthetic web crawl (~49% phase 1 / ~38% phase 2: crawl-medium/high, syn-crawl-medium/high), code 14%, math 6.4%, multilingual 5% (11 languages), nemotron-cc-code 2.1%, finepdfs (upweighted in phase 2), wiki, academic, legal, Crawl++, and synthetic SFT-style data (sft-code/stem/general); new datasets added since Super: Nemotron-Pretraining-Code-v3 (173B fresh GitHub tokens, cut-off Sep 30 2025), benchmark-seeded synthetic Multiple-Choice and Generative QA, Fact-Seeking (from Finewiki), Moral-Scenarios (CoT via Qwen3-235B), and a Legal family (regulations, case-law summaries, CaseHOLD/ContractNLI reformats, synthetic outcome/definition/diversity questions)[23]
Curriculum / stagingTwo-phase pretraining (15T diversity-biased then 5T quality-biased, transition at ~75% of pretraining) under a Warmup-Stable-Decay LR schedule; offline checkpoint merging (sliding 500B-token merge window, 25B checkpoint interval) for evaluation; final checkpoint selected over a large merge sweep; followed by the 1M LC-Phase CPT (33B tokens)[24]
Optimizernot disclosed
LR scheduleWarmup-Stable-Decay (WSD): 200B-token warmup to peak 2.5e-4, minus-sqrt decay over the final 5T tokens to 2.5e-6; MTP loss scaling factor 0.1 (0.05 per MTP block)[25]
Batch schedulenot disclosed
PrecisionNVFP4 pre-training (E2M1 datatype with two-dimensional block quantization on weights, Random Hadamard Transforms on wgrad inputs, stochastic rounding on gradients); the final 15% of the network (16 layers), Mamba output projections, latent projections, QKV and attention projections, MTP layers, and embedding layers kept in higher precision; claimed largest-scale stable NVFP4 training to date[26]
ParallelismLC-Phase: 32-way context parallelism, 8-way tensor parallelism, 128-way expert parallelism, 2-way pipeline parallelism; post-training RL on GB200 with Megatron EP groups co-located in one NVLink domain (NVLink-aware rank placement)[27]
HardwareGB200 GPUs (LC-Phase 'train on GB200 GPUs'; production RL cluster on GB200 NVL72 nodes)[27]
Disclosed computenot disclosed

Post-training

SFTyes[28]
Preference optimisationGRPO[29]
Reasoning trainingUnified RLVR across terminal usage, office/productivity, software engineering, search, tool calling, math, code, STEM, safety, chat, instruction following, long-context QA, and reasoning environments; three reasoning modes (reasoning-off, regular, medium-effort) with inference-time budget control (medium-effort ~2.5× fewer tokens at ~7% accuracy cost); >10 domain-specialized teachers; test-time scaling via generate-verify-refine on Olympiad math (IMO-ProofBench Advanced 82.3%)[30]
DistillationMulti-teacher On-Policy Distillation (MOPD): >10 domain-specialized teachers (STEM/general reasoning, competitive coding, SWE, office & workplace, search, terminal-use, conversational tool-use, usability, agentic safety, chat, instruction-following & factuality) consolidated into the student over two iterative cycles via asynchronous dense token-level supervision (negative reverse-KL on student-generated rollouts, PPO-style clipping, IcePop masking), with a light SFT warmup to align student rollouts with teacher distributions; MTP Boosting additionally aligns the draft head to backbone logits via temperature-scaled forward-KL[31]

Modality

Typetext-only
Attachmentn/a[32]

Inference efficiency

KV-cache designhybrid cache: Mamba-2 layers keep a constant-size SSM state per sequence (FP32 at train, FP16 with stochastic rounding at deploy; block-scaled INT8 explored); global-attention layers keep a small GQA KV cache (2 KV heads, FP8 at deploy); the 32-bit Mamba cache is larger than the FP8 KV cache at sequence lengths up to 64K[33]
Quantisation shippedSingle NVFP4 checkpoint at 5.03 bits-per-element: NVFP4 routed-expert GEMMs (E2M1, dynamic max-based activation scaling, Four-Over-Six 4/6 weight scaling), FP8 per-tensor shared experts and Mamba linear layers, BF16 attention linear layers / MoE latent projections / embeddings / MTP layers, FP8 KV cache, FP16-SR Mamba cache; runs native W4A4 on Blackwell and W4A16 on Hopper[34]
Speculative decodingNative Multi-Token Prediction (MTP): two shared-weight MTP blocks (each a single attention layer + single MoE layer) trained throughout pretraining/SFT/RL as an internal drafter; MTP Boosting (head-only forward-KL distillation, temperature 2, 7 MTP steps) fixes the train-inference mismatch; draft length is a deployment knob (peak 2.89× decode speedup at DL=6); SPEED-Bench average acceptance length 4.584 (greedy, boosted)[35]
Serving optimisationsTRT-LLM for the headline throughput numbers (vLLM for baselines); wide EP for high-throughput / wide TP for low-latency serving on GB200 NVL72 (single NVLink domain); prefill-decode disaggregation for hybrid Mamba-Attention landed in vLLM (~10% throughput gain); FlashInfer NVLinkOneSided all-to-all backend (~5% gain) and DWDP weight-prefetch alternative; EPLB hot-expert replication; MoE-side prefill chunking; SSM-state snapshotting at every draft step for speculative rollback and coarse-grained prefix caching[36]

Evaluation

Benchmarks (report's own numbers only)

BenchmarkValueRef
Terminal Bench 2.156.4Table 10
SWE-Bench Verified70.7Table 10
GDPVal46.7Table 10
TauBench V3 (Average)70.9Table 10
BrowseComp44.4Table 10
PinchBench90.0Table 10
ProfBench (Search)56.0Table 10
IOI 2025 (Score / 600)570.0Table 10
LiveCodeBench (v6)89.0Table 10
IMOAnswerBench (with tools)92.3Table 10
GPQA (no tools)87.0Table 10
MMLU-Pro86.8Table 10
HLE (no tools)26.7Table 10
AA-LCR65.4Table 10
RULER (1M)94.7Table 10
MMLU-ProX (avg 10 langs)83.0Table 10
IFBench (prompt loose)81.7Table 10
MMLU (base, 5-shot acc)89.08Table 2
RULER 64K (base)95.3Table 2
IMO-ProofBench Advanced (test-time scaling)82.3% (173/210)Table 11

Provenance references

  1. stated · Title, Abstract
  2. stated · Title page byline
  3. stated · Title page (first line of the PDF: '2026-6-9')
  4. stated · pipeline
  5. stated · Abstract: 'We open-source the base, post-trained, and quantized checkpoints, along with the training data and recipe on HuggingFace'; §1 (checkpoint list: Base BF16, Post-Trained BF16, NVFP4, GenRM)
  6. stated · Abstract, Table 1, §6 (550B total)
  7. stated · Abstract, Table 1, §6 (55B active per token)
  8. stated · Table 1 (Total Layers)
  9. stated · Table 1 (Model Dimension)
  10. stated · Table 1 (Expert Hidden Dimension); note: per-expert hidden dim for LatentMoE routed experts; shared expert intermediate size is 10240; MoE latent size is 2048
  11. stated · Table 1 (Q-Heads n_q)
  12. stated · Table 1 (KV-Heads n_kv) — GQA with 2 KV heads on the global attention layers
  13. stated · Table 1 (Head Dimension)
  14. stated · §2.1, §5 ('hybrid Mamba-2 stack with sparse global Attention anchors'); Table 1 (Q-Heads 64, KV-Heads 2, Head Dim 128 — the attention layers are GQA)
  15. stated · Figure 2, §2.1, §5; note: the PDF text extraction of Figure 2 is partially garbled ('x3/x4/x2' multipliers), so exact attention-layer count is not recoverable from the stored text
  16. stated · Table 1 (Total Experts per Layer)
  17. stated · Table 1 (Top-k / Activated Experts)
  18. stated · §2.5 (CPT on 1,048,576 (1M) context length for 92% of iterations)
  19. stated · §2.5 (LC-Phase CPT at 1,048,576 (1M) context for 92% of iterations, 4,096 for the remaining 8%; 33B tokens total)
  20. stated · Abstract, §3.7.2 ('supports contexts of up to 1M tokens')
  21. stated · §2.5
  22. stated · Abstract, §2.4 ('total horizon of 20 trillion tokens')
  23. stated · §2.3, Figure 4
  24. stated · §2.3.6, §2.4, §2.5
  25. stated · §2.4
  26. stated · §2.2
  27. stated · §2.5, §3.6.2
  28. stated · §3.1 (two-stage SFT: stage 1 packed 294,912-token sequences, global batch 64, 204,800 samples, cosine LR peak 1.5e-5; stage 2 packed 515,000-token sequences with long-context data up to 512K, 19,200 samples; shared-weight MTP objective retained)
  29. stated · §3.2 (unified RLVR 'largely follows the asynchronous GRPO algorithm with the stability optimizations proposed in NVIDIA (2026)'; global batch 8192, 16 rollouts/sample, max generation 48K→64K tokens); §3.3.2 (Chat teacher additionally uses RLHF with an Ultra-based generative reward model / GenRM)
  30. stated · §3.2, §3.3, §3.5, §3.7.3
  31. stated · §3.3, §3.3.1, §3.4
  32. n/a · modality.type = text-only
  33. stated · §4.5, Table 12, §5.1
  34. stated · §4.1, §4.3, §4.6, Table 12
  35. stated · §2.1, §3.4, Table 6, §5.1
  36. stated · §5.2, §5.3