NVIDIA · usage rank #11
NVIDIA-Nemotron-3-Ultra-550B-A55B
What NVIDIA-Nemotron-3-Ultra-550B-A55B’s own documents say it is built from — every method with the sentence that describes it.
- Parameters
- 550B / 55B active
- Context
- 1M
- Attention
- Mamba hybridMamba-2 layers + sparse GQA attention layers
- Experts
- LatentMoE, 512 routed + 1 shared, top-22
- Input
- text
- Licence
- OpenMDW-1.1
- Released
- 2026-06-04
Rank by open-weight tokens on OpenRouter, week of 2026-10-26.
Against the consensus recipe
Its documents state 11 of the 18 methods the field agrees on.
Data curation
- Benchmark decontaminationnot in its documents
- Synthetic data generationused
- Corpus deduplicationused
Training objective
- Multi-Token Predictionnot in its documents
- Knowledge insulationnot in its documents
- Quantization-aware distillationnot in its documents
Model architecture
- Mixture of expertscore
- Multi-Token Predictioncore
- Grouped-query attentionnot in its documents
Optimization
- Muonnot in its documents
- Expert parallelismused
- Quantization-Aware Trainingnot in its documents
Post-training
Inference & serving
- Thinking modeused
- Multi-Token Predictionused
- FP8used
Checked in its code
Its code shows 7 of the 19 architecture features checked — 3 of them are not stated in its documents.
Sources say what a lab chose to describe; the modeling code says what the checkpoint runs. Each feature below was put to a code verifier, and only the config values and source lines it cited that were found in the files are shown, linked to the line.
Read from Hugging Face Transformers (modeling_nemotron_h.py, configuration_nemotron_h.py) and config.json at revision 77df655d5e, by anthropic/claude-sonnet-5 on 2026-11-07.
| Feature | Code | Its documents | Verdict |
|---|---|---|---|
| Grouped-query attention code only | in its code | not stated | Full-attention layers use 64 query heads but only 2 key/value heads, with repeat_kv expanding the KV heads to match query groups. show evidencenum_attention_heads = 64num_key_value_heads = 2 |
| Multi-head latent attention | not in its code | not stated | The attention module is a standard multi-head/GQA attention with separate q/k/v projections, no latent KV down-projection or decoupled… show evidenceThe attention module is a standard multi-head/GQA attention with separate q/k/v projections, no latent KV down-projection or decoupled RoPE key as in DeepSeek-V2 MLA. |
| Sliding-window attention | not in its code | not stated | sliding_window is null in this config and no windowed masking code is used for the full-attention layers. show evidencesliding_window = null |
| Interleaved sliding-window and global attention | not in its code | not stated | There is no sliding-window layer type in layers_block_type (only mamba, moe, attention), so no interleaving of local/global attention… show evidenceThere is no sliding-window layer type in layers_block_type (only mamba, moe, attention), so no interleaving of local/global attention occurs. sliding_window = null |
| Indexer-selected sparse attention (DeepSeek Sparse Attention) | not in its code | not stated | No indexer module or top-k token selection mechanism is present in the attention or mixer code; config has no… show the whole verdictNo indexer module or top-k token selection mechanism is present in the attention or mixer code; config has no index_topk/index_n_heads/index_head_dim keys. |
| Linear-attention or state-space layers alongside full attention code only | in its code | not stated | layers_block_type mixes mamba (linear_attention) layers with full_attention and moe layers, dispatched via MIXER_TYPES based on block type. show evidencelayers_block_type = ["mamba", "moe", "mamba", "moe", "mamba", "moe", "mamba", "attention", "moe",… |
| Gated DeltaNet layers | not in its code | not stated | The recurrent mixer is a Mamba-2 SSM block (selective state-space with conv1d), not a gated delta-rule recurrence; no delta-rule write or… show the whole verdictThe recurrent mixer is a Mamba-2 SSM block (selective state-space with conv1d), not a gated delta-rule recurrence; no delta-rule write or Kimi Delta Attention code is present. |
| Mamba-2 layers | in its code | adopted | The linear-attention layers are implemented as NemotronHMamba2Mixer, a Mamba-2 selective state-space block with chunked SSD scan and… show evidenceThe linear-attention layers are implemented as NemotronHMamba2Mixer, a Mamba-2 selective state-space block with chunked SSD scan and per-head scalar decay (A_log/D), configured via mamba_num_heads and ssm_state_size. mamba_num_heads = 256ssm_state_size = 128 |
| Learnable attention sink | not in its code | not stated | No learned per-head sink logit is added to attention weights in eager_attention_forward or elsewhere; the softmax denominator is computed… show evidenceNo learned per-head sink logit is added to attention weights in eager_attention_forward or elsewhere; the softmax denominator is computed over only the key positions. |
| Gated attention output | not in its code | not stated | NemotronHAttention's output goes directly through o_proj with no sigmoid gate multiplying the attention output; gating (silu) is only used… show evidenceNemotronHAttention's output goes directly through o_proj with no sigmoid gate multiplying the attention output; gating (silu) is only used inside the Mamba2 mixer's RMSNormGated, not attention. |
| QK normalization | not in its code | not stated | NemotronHAttention projects q/k/v and applies no RMSNorm/LayerNorm to queries or keys before the dot product. show evidence |
| Partial RoPE | not in its code | not stated | partial_rotary_factor is 1.0, meaning RoPE is applied to the full head dimension, and no decoupled MLA-style rope key exists. show evidencepartial_rotary_factor = 1.0 |
| YaRN RoPE scaling | code unclear | not stated | No rope_scaling or rope_parameters key is present in the config, and the attention module doesn't reference rotary embeddings or scaling… show the whole verdictNo rope_scaling or rope_parameters key is present in the config, and the attention module doesn't reference rotary embeddings or scaling at all, so YaRN cannot be confirmed or ruled out from the given files. |
| Layers without positional encoding (NoPE) | code unclear | not stated | NemotronHAttention never calls apply_rotary_pos_emb or references rotary embeddings in its forward, and no rotary embedding module is… show evidenceNemotronHAttention never calls apply_rotary_pos_emb or references rotary embeddings in its forward, and no rotary embedding module is instantiated in the given files, so whether full_attention layers receive any positional encoding cannot be determined from this code. |
| Mixture of experts | in its code | adopted | MoE blocks route tokens to a subset of 512 routed experts chosen per-token by NemotronHTopkRouter with top-k of 22. show evidencen_routed_experts = 512num_experts_per_tok = 22 |
| Shared expert | in its code | adopted | Every MoE layer additionally runs a shared expert MLP of intermediate size 10240 that all tokens pass through, added to the routed-expert… show evidenceEvery MoE layer additionally runs a shared expert MLP of intermediate size 10240 that all tokens pass through, added to the routed-expert output. n_shared_experts = 1moe_shared_expert_intermediate_size = 10240 |
| Auxiliary-loss-free load balancing (selection bias) code only | in its code | not stated | The router adds e_score_correction_bias to the sigmoid scores only for top-k expert selection, not for weighting outputs (weights are… show evidenceThe router adds e_score_correction_bias to the sigmoid scores only for top-k expert selection, not for weighting outputs (weights are gathered from the unbiased scores). |
| Multi-token prediction layers | in its code | adopted | The config declares num_nextn_predict_layers=1 with an mtp_layers_block_type, and the modeling code explicitly ignores mtp.* weights on… show evidenceThe config declares num_nextn_predict_layers=1 with an mtp_layers_block_type, and the modeling code explicitly ignores mtp.* weights on load, indicating MTP modules are declared by the checkpoint. num_nextn_predict_layers = 1mtp_layers_block_type = ["attention", "moe"] |
| SwiGLU feed-forward | not in its code | not stated | Both the dense MLP and the MoE experts use a single up_proj/down_proj pair with a relu2 activation (no gating projection, no SiLU), so… show evidenceBoth the dense MLP and the MoE experts use a single up_proj/down_proj pair with a relu2 activation (no gating projection, no SiLU), so this is not SwiGLU. mlp_hidden_act = "relu2" |
token mixer 7
softmax attention 1
“select Attention layers”
sparse attention 1
fixed-pattern sparse attention 1
“a hybrid Mamba-2 stack with sparse global Attention anchors”
linear attention & state space 3
Mamba 3
“a hybrid Mamba-2 stack with sparse global Attention anchors”
“Nemotron 3 Ultra : Open, Efficient Mixture-of-Experts Hybrid Mamba-Transformer Model for Agentic Reasoning”
“--mamba-ssm-cache-dtype float16”
hybrid layer stacking 2
“Open, Efficient Mixture-of-Experts Hybrid Mamba-Transformer Model for Agentic Reasoning”
“employing a Mixture-of-Experts hybrid Mamba-Attention architecture”
channel mixer 5
mixture of experts 5
“Nemotron 3 Ultra : Open, Efficient Mixture-of-Experts Hybrid Mamba-Transformer Model for Agentic Reasoning”
expert load balancing 1
“replicating hot experts across EP ranks (EPLB)”
latent mixture of experts 2
“LatentMoE (Elango et al., 2026), which buys more routed experts at fixed inference cost”
“introduce a breakthrough hybrid latent mixture-of-experts (MoE) architecture”
prediction head 5
“the Ultra model incorporates Multi-Token Prediction (MTP) layers for faster text generation and improved quality”
“Multi Token Prediction (MTP) leads to faster inference with speculative decoding”
“The model includes Multi-Token Prediction (MTP) layers using a shared-weight design across prediction heads”
“the hidden states passed as input to the MTP at step k are sampled from the set of hidden states produced at MTP steps 1,..,k−1”
“--speculative-config '{"method": "nemotron_h_mtp", "num_speculative_tokens": 5}'”
1 more mentioned or evaluated
expert load balancing “To quantify expert imbalance, we measure the MaxVio metric (DeepSeek-AI, 2025b) which calculates the peak load on any single expert…”
Not used 1
fine-grained experts “LatentMoE helped us achieve better accuracy per parameter than standard Granular MoEs”
language modelling objective 1
“the sampled-token objective applies supervision only to realized actions”
multi-token prediction objective 3
“The goal of MTP Boosting is to make the MTP head match the backbone’s next-token distribution under the input conditions or noise it…”
“The backbone is fixed for the entire training phase, and only the MTP head receives gradient updates.”
“As in pre-training, we retain the shared-weight MTP objective during SFT, using two MTP layers with a per-token auxiliary-loss scaling…”
distillation objective 1
“We use a temperature-scaled forward-KL loss against the backbone’s logits.”
auxiliary loss 1
“We mask the loss on unfinished trajectories”
1 more mentioned or evaluated
distillation objective “A natural alternative to the sampled-token objective is distribution-level distillation, where the student is trained to match the…”
optimization 28
learning-rate schedule 3
“using a cosine learning-rate schedule with peak learning rate 1.5×10−5”
“Warmup-Stable-Decay (WSD) learning rate schedule over a total horizon of 20 trillion tokens”
“starting learning rate annealing (both a 5T and 10T decay) immediately after rolling back to the 15T token checkpoint mitigates divergence”
training precision 8
“The majority of linear layers use NVFP4 for weights, activations, and gradients”
“it is trained using an NVFP4 pre-training recipe to maximize compute efficiency”
“BF16 precision for attention linear layers, MoE latent projection layers”
“NVFP4 layers use the E2M1 datatype”
“Rolling back to an earlier checkpoint and moving back to the full FP32 gradient reduction recipe re-stabilized training”
“this step is precisely the introduction of the mixed-FP8 layers on top of the NVFP4-amax recipe”
“select layers (including latent projections, MTP layers, QKV/attention projections, and embeddings) are maintained in BF16 or MXFP8 for…”
“NVFP4 layers use the E2M1 datatype with two-dimensional block quantization on weights”
quantization-aware training 1
“FP16 with stochastic rounding preserves FP32-cache accuracy”
training stability 1
“the route we have taken is to pad the affected weight matrices at load time”
training parallelism 4
“wide EP the practical choice for high-throughput serving”
“wide TP the practical choice for low-latency serving”
“32-way context parallelism”
“2-way pipeline parallelism”
training runtime 9
“offline checkpoint merging for evaluation analysis”
“Enabling asynchronous checkpointing from the Nvidia Resiliency Extension (NVRx), where model parameters are copied to CPU and persisted in…”
“At job completion, a single designated sidecar process archives the caches back to shared storage as compressed tarballs, rather than all…”
“we cache the distributed save plan so it is computed once rather than on every save.”
“leveraging in-flight weight updates and MTP to accelerate rollout generation”
“overlapped NCCL transfers with D2H copies to shrink the synchronous staging window”
“persistent checkpoint worker processes that avoid fork/spawn overhead on each save”
“Persistent warm cache on shared storage.At job completion, all JIT artifacts ... are written back to a persistent shared directory as…”
“Before Ray initialization, a setup script on each node extracts the warm cache tarballs into node-local/tmp”
1 more mentioned or evaluated
training parallelism “DWDP (nvidia2026dwdp) keeps execution data-parallel and instead pulls each layer's expert weights to every rank”
Not used 1
training precision “The first divergence, which occurred at around 8T tokens, was attributed to a reduction in local gradient accumulation precision for the…”
filed at the root 1
“The data was subsequently deduplicated and filtered in various ways described in the Nemotron-CC paper”
data sourcing 7
“The GitHub Crawl was collected using the GitHub REST API and the Amazon S3 API”
“The English Common Crawl data was downloaded from the Common Crawl Foundation”
“These datasets, such as FinePDFs, EssentialWeb, HotpotQA, SQuAD, and HelpSteer3”
“These datasets, such as FinePDFs, EssentialWeb, HotpotQA, SQuAD, and HelpSteer3”
“The foundation of the model is trained on the Nemotron-3-Ultra corpus”
“Nemotron-CC paper”
“Major portions of the fine-tuning corpus are released in the Nemotron-Post-Training-v3 collection”
data filtering 5
“we applied just heuristic filtering instead—similar to what we did for lower quality English data in the Nemotron-CC pipeline”
“we apply targeted keyword- and regex-based filters and remove all trajectories matching such behavior”
“We then aggressively filter reasoning traces exhibiting pathological repetition, such as repeated n-grams within a sliding window or…”
“Across all pipelines we stack automated verification—compilers, numerical checks, language identification—to ensure our data is high quality”
“we apply a unified data filtering pipeline to ensure that only high-quality, license-compliant, and verifiable samples are used for…”
deduplication 2
“The data was subsequently deduplicated and filtered in various ways”
“we enforce an in-pack deduplication constraint to prevent identical prompts from co-occurring within the same sequence.”
synthetic data 2
“we generate synthetic data by distilling trajectories, solutions, and translations from strong teacher models”
“During post-training, we generate synthetic data by distilling trajectories, solutions, and translations from strong teacher models”
data mixture & curriculum 2
“For data mixture and curriculum construction, we adopt the Gaussian-based approach introduced in NVIDIA (2025b).”
“two-phase curriculum and transition from a data mixture which biases dataset diversity ... to data mixture which biases dataset quality”
sequence packing 1
“we adopt a length-aware best-fit packing strategy (Ding et al., 2024), which packs multiple conversations into sequences up to a maximum…”
1 more mentioned or evaluated
synthetic data “mitigation strategies like counterfactual data augmentation to align with the desired model behavior”
supervised fine-tuning 4
“the student undergoes a very light SFT on data drawn from the teacher’s training distribution.”
“To train the student model, we start from the pre-trained base model and perform Supervised Fine-Tuning (SFT) in two stages”
“rejection sampling on those reasoning traces”
“We construct the final SFT mixture using token-level target proportions rather than example counts”
reinforcement learning algorithm 9
“post-trained using Supervised Fine Tuning (SFT), Reinforcement Learning (RL), and Multi-teacher On-Policy Distillation”
“Our training procedure largely follows the asynchronous GRPO algorithm with the stability optimizations proposed in NVIDIA (2026)”
“we conduct a unified RLVR (Reinforcement Learning with Verifiable Reward) training stage spanning all available environments”
“the teacher also underwent abstention training, where the model was encouraged to abstain when uncertain rather than hallucinate incorrect…”
“where represents token-level masking using IcePop strategy (ring1t)”
“We dynamically calibrated the abstention reward throughout training”
“we also penalize malformed reasoning and tool calls by assigning negative advantage to the offending tokens.”
“proceeded with pivot RL in the MOPD stage, distilling the SFT-trained teacher into the student Ultra model”
“PPO-style clipping is applied to r_t(θ) around the proximal policy πprox.”
reward modelling 3
“Generative Reward Models (GenRM) (Wang2025HelpSteer3Preference) with reasoning capabilities help mitigate such reward hacking behaviors”
“length-based adjustments are applied to the RL rewards for them.”
“A principle-following GenRM makes this process much more flexible”
preference optimization 1
“Conversational quality was further refined through RLHF”
policy distillation 4
“MOPD provides a dense token-level learning signal from the relevant teacher distribution.”
“In our implementation, MOPD is executed asynchronously.”
“To stabilize this setting, we decouple the behavior policy from the proximal policy”
“m_t represents token-level masking using IcePop strategy”
rollout & RL infrastructure 3
“During RL and MOPD, we train using a one-step off-policy asynchronous RL setup, so rollout generation is overlapped with the policy update”
“It utilized an asynchronous RL architecture that fully decouples training from inference across separate GPU devices”
“This technique uses strong teacher models to guide training on the model's own generated attempts (on-policy rollouts)”
agentic post-training 2
“unified RLVR over a wide mix of reasoning, agentic, code, safety, usability, and chat environments”
“advanced reinforcement learning techniques with concurrent multi-environment post-training at scale”
mid-training & continual pretraining 2
“added a long-context phase (LC-Phase) at the end of pretraining”
“continuous pretraining (CPT) to equip the base model with long-context ability”
Not used 1
policy distillation “yielding stronger gains than purely off-policy distillation”
decoding strategy 5
“At each decoding iteration, the MTP head is applied recurrently to propose 𝑘 candidate tokens, which the base model verifies in a single…”
“enables faster inference via native speculative decoding”
“In order to accelerate rollout generation, we use speculative decoding with Multi-Token Prediction (MTP).”
“speculative-algorithm EAGLE”
“best-of-n selection traces”
reasoning control 6
“Ultra is also equipped with reasoning effort control, which supports inference-time adjustment of the accuracy–compute trade-off”
“generate both COT and TIR solution trajectories with DeepSeek-V4-Pro”
“Following the generate–verify–refine methodology of Shao et al. (2025) and Nemotron-Cascade-2 (Yang et al., 2026)”
“Each model runs Terminal-Bench v2.1 at four budgets - 10, 20, 50 and 100 turns”
“we customized the Terminus 2 harness so that every run states the turn budget up front... and warns the model as it approaches the limit”
“The regular and medium-effort reasoning modes have the option to be used in conjunction with inference-time budget control.”
KV cache management 2
“All deployment snippets below default to port 8000, with chunked prefill, NVFP4 KV caching, and MTP (5 speculative tokens) enabled”
“enable-prefix-caching”
inference quantization 14
“FP8 per-tensor GEMMs for shared experts and Mamba linear layers; static max-calibrated per-tensor scales.”
“We therefore selected5.03 BPE(NVFP4 with mixed-FP8) as the operating point”
“NVFP4 routed-expert GEMMs; Dynamic max-based activation scaling and max-calibrated 4/6 weight scaling.”
“BF16 precision for attention linear layers, MoE latent projection layers (i.e, layers for which accuracy degradation outweighed the…”
“All deployment snippets below default to port 8000, with chunked prefill, NVFP4 KV caching, and MTP (5 speculative tokens) enabled”
“KV Cache is quantized to FP8”
“KV Cache is quantized to FP8”
“We apply post-training quantization (PTQ) using Model-Optimizer to quantize the Nemotron 3 Ultra checkpoint to NVFP4”
“Four-Over-Six selects between the 𝑀 = 4 and 𝑀 = 6 weight grids per block, using the option that minimizes reconstruction error.”
“we quantize the Mamba SSM cache from the original FP32 precision to lower precisions”
“we use FP16 SSM cache storage with stochastic rounding”
“We begin from a heuristic mixed per-layer precision recipe informed by Model-Optimizer AutoQuantize sensitivity analysis”
“Random Hadamard Transforms on inputs to wgrad”
“W4A16 (weights NVFP4, activations BF16)”
serving parallelism 4
“Prefill-decode disaggregation, which runs the two phases on separate workers with their own parallelism and scheduling, is the established…”
“The fix was to make rank assignment domain-aware and ensure all GPUs within an EP group are co-located on the same rack.”
“enable_attention_dp: true”
“use_low_precision_moe_combine: true”
inference scheduling 3
“The fix was to explicitly bind policy and vLLM worker processes to the CPU socket local to their assigned GPUs.”
“bundle indices are sorted by a( domain_min_topo_rank,topo_rank,gpu_id )composite key”
“and registers it as a Ray custom resource (nvlink_domain_<ClusterUUID>), giving the scheduler rack-membership information it previously…”
inference kernel 2
“The fix is MoE-side chunking, splitting the token batch within the MoE kernel itself”
“Marlin NVFP4 linear and MoE kernels on Hopper that impose tile- and thread-level alignment constraints”
context management 1
“We focus primarily on discard-all context management, where earlier search observations are removed once the interaction history exceeds…”
agentic scaffolding 2
“we install a runtime command filter that blocks every way the agent could pull that history back over the network”
“generate both COT and TIR solution trajectories”
9 more mentioned or evaluated
inference quantization “MSE calibration gave an additional weight-MSE reduction of 27.1%, downstream evaluations did not show a consistent accuracy increase”
KV cache management “We ran initial experiments on Nemotron 3 Super NVFP4, using emulated quantization”
inference quantization “Block-scaled INT8 quantization with stochastic rounding largely preserves FP32-cache accuracy”
inference quantization “FP8 E4M3 quantization degrades accuracy”
inference quantization “we experimented with max-based, MSE-based, and Four-Over-Six scaling”
KV cache management “we also explore 8-bit Mamba cache quantization formats. We evaluated this on Nemotron 3 Super.”
inference quantization “We selected the model’s bits-per-element (BPE) budget empirically, by quantizing a fixed intermediate checkpoint at a range of BPE…”
context management “including discard-all resets and summary-based compression.”
reasoning control “There are a range of approaches to this today including model ‘effort’ parameters, token limits at the request or task level, or turn…”
filed at the root 2
“Data Designer is one of the libraries used to prepare these corpora”
“software frameworks (e.g., CUDA libraries)”
inference engine 4
“disaggregation for hybrid Mamba-Attention models now works out-of-the-box in vLLM (Kwon et al., 2023)”
“Library paths and environment variables were explicitly forwarded to all subprocesses at spawn time.”
“SGLang deployment snippet provided”
“VLLM_FLASHINFER_MOE_BACKEND=latency(TRTLLM-Gen)”
training framework 4
“We used transformers on a single node with CPU offloading to quantize Ultra layer-by-layer.”
“we chose Megatron-LM for its efficient multi-node distributed parallelism and MoE support”
“The RL environments and datasets are released as part of NeMo Gym”
“Runtime Engine(s): NeMo 26.04.01”
kernel & quantization library 6
“mamba-backend flashinfer”
“adopted FlashInfer’s NVLinkOneSided implementation... as the best end-to-end choice for Nemotron 3 Ultra serving on GB200 NVL72”
“Model-Optimizer supports both HuggingFace and Megatron-LM for quantization”
“precompiled cubins were baked into the container image at build time viaflashinfer download-cubin”
“moe_config: backend: CUTEDSL”
“VLLM_FLASHINFER_MOE_BACKEND=throughput(CUTLASS)”
agent product 4
“the model is trained under at least two of the following harnesses: • Stirrup • OpenHands • OpenCode • Terminus • Droid”
“the model is trained under at least two of the following harnesses: • Stirrup • OpenHands • OpenCode • Terminus • Droid”
“the model is trained under at least two of the following harnesses: • Stirrup • OpenHands • OpenCode • Terminus • Droid”
“the model is trained under at least two of the following harnesses: • Stirrup • OpenHands • OpenCode • Terminus • Droid”
infrastructure service 6
“All components of the RL stack were aligned to a single version of shared GPU kernel libraries, ensuring ABI consistency across the job.”
“we rewrite the in-container repository to look like a fresh clone taken at the moment of the task’s base commit”
“Enroot’s local squashfs cache ensures that once a node has extracted the container image, subsequent jobs reuse the cached local copy”
“We eliminated 40% of actor registrations by converting short-lived actors to tasks, pooling initialization actors per node, and applied…”
“We eliminated 40% of actor registrations by converting short-lived actors to tasks, pooling initialization actors per node, and applied…”
“The recommended multi-processing backend for multi-node deployments is Ray v2”
evaluation 15
benchmark 8
“BrowseComp with Search, Tau Bench 3, ProfBench with Search, PinchBench, Vals.ai, LongBench v2”
“the evaluations also used dedicated open-source packaged containers for ScaleAI Multi Challenge Multi Turn Instruction Following and…”
“BrowseComp with Search, Tau Bench 3, ProfBench with Search, PinchBench, Vals.ai, LongBench v2”
“BrowseComp with Search, Tau Bench 3, ProfBench with Search, PinchBench, Vals.ai, LongBench v2”
“BrowseComp with Search, Tau Bench 3, ProfBench with Search, PinchBench, Vals.ai, LongBench v2”
“the evaluations also used dedicated open-source packaged containers for ScaleAI Multi Challenge Multi Turn Instruction Following and…”
“These datasets, such as FinePDFs, EssentialWeb, HotpotQA, SQuAD, and HelpSteer3”
“BrowseComp with Search, Tau Bench 3, ProfBench with Search, PinchBench, Vals.ai, LongBench v2”
evaluation harness 3
“All evaluation results were collected via Nemo Evaluator SDK”
“We used three main evaluation harnesses: Nemo Gym, Nemo Skills, and Harbor”
“We used three main evaluation harnesses: Nemo Gym, Nemo Skills, and Harbor”
judge 2
“a deterministic verifier marks an injection as resisted only if the agent does not invoke the attacker’s target tool with the target…”
“We then check each trajectory against the reference answer using an LLM-based judging pipeline”
human & real-world evaluation 2
“These benchmarks were reserved for final validation: they were not used for training-time monitoring, checkpoint selection, or any other…”
“we use an automated red-teaming loop in which an attacker model iteratively rewrites the injected instruction against a defender model”
unfiled 12
“Nemotron 3 Ultra : Open, Efficient Mixture-of-Experts Hybrid Mamba-Transformer Model for Agentic Reasoning”
“extended sandboxing support via AWS ECS on Nemo Evaluator”
“use open source permissive models such as GPT-OSS-120B to produce step-by-step reasoning traces”
“These datasets, such as FinePDFs, EssentialWeb, HotpotQA, SQuAD, and HelpSteer3”
“These datasets, such as FinePDFs, EssentialWeb, HotpotQA, SQuAD, and HelpSteer3”
“The fix was a three-pronged cache management strategy”
“The trial service is governed by the NVIDIA API Trial Terms of Service”
“Use of this model is governed by the OpenMDW License Agreement, version 1.1 (OpenMDW-1.1)”
“Restructuring to a single multi-node srun reduced Slurm controller interactions to 𝑂(1).”
“BrowseComp with Search, Tau Bench 3, ProfBench with Search, PinchBench, Vals.ai, LongBench v2”
“the model is trained under at least two of the following harnesses: • Stirrup • OpenHands • OpenCode • Terminus • Droid”
“The SWE teacher was trained through a three-stage pipeline.”
Further reading
Picked by hand, not extracted: where to read more, not evidence for anything on this page.
- NVIDIA-Nemotron-3-Ultra-550B-A55B architecture card — LLM Architecture Gallery explainer sebastianraschka.com
Sources
The curated source list for this model. "not fetched" means the URL is recorded but its text was not read in the current run.
| Source | Kind | Publisher | Status |
|---|---|---|---|
| research.nvidia.com/labs/nemotron/files/NVIDIA-Nemotron-3-Ultra-Technic… | technical report | NVIDIA | read |
| arxiv.org/abs/2606.15007 | technical report | NVIDIA | read |
| research.nvidia.com/labs/nemotron/Nemotron-3-Ultra | official blog | NVIDIA Research | read |
| nvidianews.nvidia.com/news/nvidia-debuts-nemotron-3-family-of-open-mode… | official blog | NVIDIA Newsroom | read |
| huggingface.co/nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B-BF16 | model card | NVIDIA | read |
| build.nvidia.com/nvidia/nemotron-3-ultra-550b-a55b/modelcard | vendor docs | NVIDIA | read |
| artificialanalysis.ai/articles/nvidia-nemotron-3-ultra-released | third party analysis | Artificial Analysis | read |