NVIDIA Nemotron™ is a family of open models with open weights, training data, and recipes, delivering leading efficiency and accuracy for building specialized AI agents.
NVIDIA-Nemotron-3.5-Lightning-30B-A3B-BF16 is a large language model (LLM) trained by NVIDIA. This is the full-precision (BF16) release of Nemotron 3.5 Lightning — the reference weights of the model, intended primarily as the starting point for customization: post-training (SFT, RL, distillation), domain adaptation, and producing your own quantized or GGUF variants. For latency- and throughput-optimized inference, use the NVFP4 release instead.
The model employs a hybrid Mixture-of-Experts architecture, utilizing interleaved Mamba-2 and MoE layers, along with select Attention layers. The Lightning 3.5 model is released alongside a number of speculative decoding methods for faster text generation. The model has 3B active parameters and 30B parameters in total.
For running Nemotron 3.5 Lightning fast — with NVFP4 quantization, W4A16 for broad hardware coverage, and the DSpark recipe for DGX Spark — please see: NVIDIA-Nemotron-3.5-Lightning-30B-A3B-NVFP4
We evaluated our model on the following benchmarks:
Task
Nemotron-3.5-Lightning-30B-A3B-BF16
Qwen 3.6 35B A3B
Gemma 4 26B A4B
Nemotron 3 Nano
Nemotron 3 Super
GPT-OSS 20B
General Knowledge
MMLU Pro
81.94
85.63
85.20
78.46
83.89
76.40
AA-Omniscience
17.50
19.47
22.17
20.15
26.68
16.62
Reasoning
GPQA Diamond (no tools)
75.44
83.40
79.61
74.05
78.60
71.46
HLE (text-only, no tools)
11.72
19.56
17.42
10.89
20.30
13.76
SciCode
32.60
35.33
40.28
30.08
35.11
38.63
Coding & Agentic
SWE-bench Verified
51.56
70.12
57.40
34.08
63.08
52.44
SWE-bench Multilingual
39.33
63.40
43.40
14.07
49.80
41.93
Terminal-Bench 2.1
24.58
44.38
37.22
8.29
39.61
15.17
PinchBench
85.37
88.07
74.70
66.11
80.36
57.20
BrowseComp
36.97
48.74
26.30
13.74
22.77
–
τ³-bench (Banking)
9.28
10.52
14.02
7.01
12.37
–
GDPval-AA-V2
832
1015
807
473
746
–
Instruction Following
IFBench (loose)
71.88
63.71
77.25
72.17
71.92
68.50
Long Context
AA-LCR
52.00
61.06
57.56
32.75
58.44
32.88
Accuracy numbers measured by NVIDIA under a consistent harness (NeMo Gym / Nemo Evaluator SDK); they may differ from vendors' self-reported numbers.
For reproducibility, the evaluation recipes, installation instructions, and commands for NVIDIA Nemotron 3.5 Lightning were collected and published in NeMo Gym. The reported results cover the release evaluation suite, including knowledge and reasoning, instruction following, coding, agentic, tool-use, and long-context. Most evaluations use NeMo Gym-native harnesses while a small subset, including SWE-Bench and Terminal-Bench, used NeMo Evaluator natively. The published recipes specify the benchmark-specific containers, prompts, inference parameters, parser configurations, and scoring settings used to produce the results.
NVIDIA-Nemotron-3.5-Lightning-30B-A3B-BF16 is the full-precision release of a general purpose reasoning and chat model, and is intended primarily for customization and post-training rather than direct production inference. It is intended to be used by developers who want to: post-train the model on their own data (SFT, RL via NeMo RL and NeMo Gym, or distillation), adapt it to a domain or task, produce quantized variants (NVFP4, W4A16, GGUF) for their own deployment targets, or run full-precision research and evaluation. English and coding languages are the primary languages, with Spanish, French, German, Italian, and Japanese also supported.
For developers who want to deploy Lightning 3.5 directly — in AI agent systems, chatbots, RAG systems, and instruction-following applications — the NVFP4 release is the recommended path, with optimized recipes for data centre and DGX Spark deployments.
The model was pre-trained with over 20T tokens and supports up to 1M context length. The pre-training phase used an NVFP4 recipe. The model includes Multi-Token Prediction (MTP) layers, which predict multiple future tokens to provide richer training signals.
NVIDIA-Nemotron-3.5-Lightning-30B-A3B-BF16 model was pre-trained using an NVFP4 recipe with crawled and synthetic code, math, science, and general knowledge data.
Stage 2: Continued Pre-Training for Multi-Token Prediction (MTP)
The model underwent a continued pre-training phase to train its Multi-Token Prediction (MTP) layers. In this stage, MTP heads learn to predict multiple future tokens, providing richer training signals to the base model. This phase aligns the MTP layers with the base model's distribution.
Stage 3: Supervised Fine-Tuning
The model was further fine-tuned on synthetic code, math, science, tool calling, instruction following, structured outputs, and general knowledge data. This stage incorporated data designed to support long-range retrieval and multi-document aggregation.
Stage 4: Reinforcement Learning
The model underwent multi-environment reinforcement learning using GRPO (Group Relative Policy Optimization) across math, code, science, instruction following, multi-step tool use, multi-turn conversations, and structured output environments. It utilized an asynchronous RL architecture that decouples training from inference and leverages MTP to accelerate rollout generation.
Other Properties Related to Input: Maximum context length up to 1M tokens. Supported languages include English, Spanish, French, German, Italian, and Japanese.
Other Properties Related to Output: Maximum context length up to 1M tokens
Our AI models are designed and/or optimized to run on NVIDIA GPU-accelerated systems. By leveraging NVIDIA's hardware (e.g. GPU cores) and software frameworks (e.g., CUDA libraries), the model achieves faster training and inference times compared to CPU-only solutions.
The integration of foundation and fine-tuned models into AI systems requires additional testing using use-case-specific data to ensure safe and effective deployment. Following the V-model methodology, iterative testing and validation at both unit and system levels are essential to mitigate risks, meet technical and functional requirements, and ensure compliance with safety and ethical standards before deployment.
These recipes serve BF16 directly for research, evaluation, and post-training. For production inference, the NVFP4 release is faster and covers more hardware. On a single H100, BF16 is memory-bound to ~256K — use 8× H100, GB200, or B200 for the full 1M window.
The BF16 recipes below cover vLLM and SGLang. For TensorRT-LLM, W4A16 (Blackwell / Hopper / Ampere from a single checkpoint), and DGX Spark recipes, see the NVFP4 card.
Lightning 3.5 ships with two external draft models for speculative decoding as well as MTP (Multi-Token Prediction). While DSpark is our recommended default, some workloads are better served without speculative decoding, or by DFlash or MTP:
DSpark: A semi-autoregressive speculative-decoding drafter that proposes a whole block of candidate tokens in a single forward pass from a parallel backbone. This is recommended for DGX Spark, as well as low-concurrency data centre deployments.
DFlash: A speculative-decoding drafter that uses a lightweight block-diffusion model to generate an entire draft block in one forward pass.
MTP: A modeling technique that trains the network to predict several future tokens at each position instead of only the next one.
For max throughput deployments, use the following configuration, no speculative decoding strategy is best for this serving configuration, and due to memory constraints the Mamba cache dtype is set as FP16:
Context Length: If you're memory-constrained — or want more KV-cache headroom at high concurrency — lower --max-model-len to match your workload and drop VLLM_ALLOW_LONG_MAX_MODEL_LEN=1.
The blocks above run the balanced baseline. To enable one of the speculative decoding strategies described above, append the matching flags to any command (the DFlash and DSpark draft models are the NVFP4 checkpoints):
MTP — the draft head is embedded in the target checkpoint (no separate download):
Context length: The commands above serve a 256K context window by default (or 1M in the B200 case). Set --context-length to a different value if you're memory-constrained or want more KV-cache headroom at higher concurrency.
The examples below use the OpenAI-compatible client and work with the serving backend above. Recommended sampling settings are Temperature 1.0 and Top_P 0.95.
The vLLM snippets above pass the checkpoint to --model $MODEL_CKPT, and vLLM registers the model under that same identifier (nvidia/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-BF16) — none of the commands set --served-model-name. For the other backends — or if you add a --served-model-name of your own — copy the identifier returned by GET /v1/models into MODEL below.
from openai import OpenAI
client = OpenAI(base_url="http://localhost:8000/v1", api_key="EMPTY")
MODEL = "nvidia/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-BF16"
Lightning 3.5 exposes reasoning control through chat-template kwargs: thinking enabled (the default), and thinking disabled for direct answers.
Reasoning ON / OFF and streaming examples: Click to expand!
Reasoning ON (default)
response = client.chat.completions.create(
model=MODEL,
messages=[{"role": "user", "content": "Write a haiku about GPUs"}],
max_tokens=16000,
temperature=1.0,
top_p=0.95,
extra_body={"chat_template_kwargs": {"enable_thinking": True}}
)
print(response.choices[0].message.content)
Reasoning OFF
response = client.chat.completions.create(
model=MODEL,
messages=[{"role": "user", "content": "What is the capital of Japan?"}],
max_tokens=16000,
temperature=1.0,
top_p=0.95,
extra_body={"chat_template_kwargs": {"enable_thinking": False}}
)
print(response.choices[0].message.content)
Streaming
stream = client.chat.completions.create(
model=MODEL,
messages=[{"role": "user", "content": "Explain speculative decoding in two sentences"}],
max_tokens=16000,
temperature=1.0,
top_p=0.95,
stream=True,
)
for chunk in stream:
print(chunk.choices[0].delta.content or "", end="", flush=True)
Data Modality: Text Training Data Size: More than 20 Trillion Tokens Dataset partition:Training [100%], testing [0%], validation [0%]Time period for training data collection: 2013 to December 2025 Time period for testing data collection: 2013 to December 2025 Time period for validation data collection: 2013 to December 2025 Data Collection Method by dataset: Hybrid: Automated, Manually-Collected, Synthetic Labeling Method by dataset: Hybrid: Automated, Manually-Labeled, Synthetic
NVIDIA-Nemotron-3.5-Lightning-30B-A3B-BF16 is pre-trained on a large corpus of high-quality curated and synthetically-generated data. It is trained in the English language, as well as 19 other spoken languages and 43 programming languages. Our sources cover a variety of document types such as: webpages, dialogue, articles, and other written materials. The corpus spans domains including legal, math, science, finance, and more. We also include a small portion of question-answering, and alignment style data to improve model accuracy. The model was pre-trained for more than 20 trillion tokens.
The post-training corpus for NVIDIA-Nemotron-3.5-Lightning-30B-A3B-BF16 consists of high-quality curated and synthetically-generated data. Primary languages used for post-training include English, French, German, Italian, Japanese, Spanish, and Chinese.
These datasets, such as FinePDFs, EssentialWeb, HotpotQA, SQuAD, and HelpSteer3, do not collectively or exhaustively represent all demographic groups (and proportionally therein). For instance, these datasets do not contain explicit mentions of demographic classes such as age, gender, or ethnicity in 64-99% of samples, depending on the source. In the subset where such terms are present, document-based datasets (FinePDFs and EssentialWeb) contain representational skews, such as references to "male" outnumbering those to "female", and mentions of "White" as the most frequent among ethnic identifiers (comprising 43-44% of ethnicity mentions). To mitigate these imbalances, we recommend considering evaluation techniques such as bias audits, fine-tuning with demographically balanced datasets, and mitigation strategies like counterfactual data augmentation to align with the desired model behavior. This evaluation used a 3,000-sample subset per dataset, identified as the optimal threshold for maximizing embedder accuracy.
During post-training, we generate synthetic data by distilling trajectories, solutions, and translations from strong teacher models and agent systems, often grounded in real tasks or documents and aggressively filtered for quality. For math, code, and science, we start from curated problem sets and use open source permissive models such as GPT-OSS-120B to produce step-by-step reasoning traces, candidate solutions, best-of-n selection traces, and verified CUDA kernels. For long-context and science, we build synthetic QA and reasoning data by retrieving passages from long documents, generating MCQ/OpenQA questions and answers, and paraphrasing them into multiple prompt/response formats to ensure diversity. Across all pipelines we stack automated verification—compilers, numerical checks, language identification—to ensure our data is high quality.
For all domains, we apply a unified data filtering pipeline to ensure that only high-quality, license-compliant, and verifiable samples are used for post-training. We first discard malformed examples using structural checks (e.g., missing tool definitions when tool calls are present). We then aggressively filter reasoning traces exhibiting pathological repetition, such as repeated n-grams within a sliding window or across the entire trajectory, which we found to be a strong indicator of malformed or low-quality reasoning. Finally, based on internal audits of synthetically generated datasets, we observed that some teacher models occasionally produce reasoning traces and final responses that implicitly align with specific political entities or promote nationalistic narratives. To mitigate this, we apply targeted keyword- and regex-based filters and remove all trajectories matching such behavior.
Alongside the model, we release our final pre-training and post-training data, as outlined in this section. For ease of analysis, there is a sample set that is ungated. For all remaining code, math and multilingual data, gating and approval is required, and the dataset is permissively licensed for model training purposes.
A massive collection of English web data filtered from Common Crawl, including 2.5T+ tokens of new organic, translated, and synthetically rephrased content.
Nemotron-CC-Code-v1
427.9B
High-quality code tokens extracted from Common Crawl using the Lynx + LLM pipeline to preserve structure and equations.
Nemotron-Pretraining-Code-v1 & v2 & v3
1.7T
Curated GitHub code references with multi-stage filtering, deduplication, and large-scale synthetic code data.
Nemotron-CC-Math-v1
133.3B
High-quality math pre-training dataset preserving LaTeX formatting and mathematical structures.
Crawled and Scraped from Online Sources by NVIDIA#
The English Common Crawl data was downloaded from the Common Crawl Foundation (see their FAQ for details on their crawling) and includes the snapshots CC-MAIN-2013-20 through CC-MAIN-2025-13. The data was subsequently deduplicated and filtered in various ways described in the Nemotron-CC paper. Additionally, we extracted data for fifteen languages from the following three Common Crawl snapshots: CC-MAIN-2024-51, CC-MAIN-2025-08, CC-MAIN-2025-18. The fifteen languages included were Arabic, Chinese, Danish, Dutch, French, German, Italian, Japanese, Korean, Polish, Portuguese, Russian, Spanish, Swedish, and Thai. As we did not have reliable multilingual model-based quality classifiers available, we applied just heuristic filtering instead—similar to what we did for lower quality English data in the Nemotron-CC pipeline, but selectively removing some filters for some languages that did not work well. Deduplication was done in the same way as for Nemotron-CC.
The GitHub Crawl was collected using the GitHub REST API and the Amazon S3 API. Each crawl was operated in accordance with the rate limits set by its respective source, either GitHub or S3. We collect raw source code and subsequently remove any having a license which does not exist in our permissive-license set.
Dataset
Modality
Dataset Size
Collection Period
Collecting Organisation
English Common Crawl
Text
3.36T
4/8/2025
NVIDIA Advanced Deep Learning Research
English Common Crawl 1.1
Text
Not disclosed
10/2/2025
NVIDIA Advanced Deep Learning Research
Multilingual Common Crawl
Text
812.7B
5/1/2025
NVIDIA Advanced Deep Learning Research
GitHub Crawl
Text
747.4B
4/29/2025
NVIDIA Advanced Deep Learning Research
GitHub Crawl 1.1
Text
172.7B
9/30/2025
NVIDIA Advanced Deep Learning Research
Private Non-publicly Accessible Datasets of Third Parties#
Dataset
Model(s) used
Global Regulation
Unknown
TAUS Translation Memory
Unknown
Scale HLE
Unknown
HackerRank Coding
Unknown
RL data for Search
Gemini 3; GPT-5
Mercor SWE-AgentsV1
Undisclosed
Private Non-publicly Accessible Datasets by NVIDIA#
Dataset
Model(s) used
Simple Minesweeper
Undisclosed
Simple Sudoku
Undisclosed
Multitool Typewriter Hard
Undisclosed
Machine Translation of News Commentary and TAUS Translation Memory
CommonPile (caselaw_access_project_filtered); California Code of Regulations; Judicial Ethics Opinions; GLOBALCIT; CUAD; Nemotron Personas; ToSDR Terms of Service Corpus; CodeHima/TOS_Dataset; ContractNLI; CaseHOLD; Code of Federal Regulations; Canadian Case Law (subsets that allow commercial use)
import os
os.environ["HF_ENDPOINT"] = "https://freeweights.allternit.com"
from transformers import AutoModelForCausalLM, AutoTokenizer
tok = AutoTokenizer.from_pretrained("nvidia/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-BF16", revision="a9904d24bcc1d289a1950fa9d2b978c47cf903b9")
model = AutoModelForCausalLM.from_pretrained("nvidia/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-BF16", revision="a9904d24bcc1d289a1950fa9d2b978c47cf903b9", device_map="auto")
nvidia <huggingface.co/nvidia>Newsgroups: alt.freeweights.nemotronSubject: 
lmstudio-community