NVIDIA-Nemotron-3.5-Lightning-30B-A3B-BF16

Model Summary

Model Overview

Model Developer: NVIDIA Corporation

Model Dates: December 2025 - May 2026

Data Freshness:

• The pre-training data has a cutoff date of September 2025.
• The post-training data has a cutoff date of May 2026.

What is Nemotron?

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.

Description

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 Latent Mixture-of-Experts (LatentMoE) 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.

This model is ready for commercial use.

Quick Start

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

To get quickly started on a single H100 you can use the following command.

Grab the model:

export MODEL_CKPT=nvidia/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-BF16
export DFLASH_CKPT=nvidia/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-NVFP4-DFlash

Run it with vLLM — this recipe uses DFlash speculative decoding for low-concurrency serving. (Nightly: vllm/vllm-openai:nightly-821717118fc26667dd474b9b0ab81d29259dfc5c)

VLLM_ALLOW_LONG_MAX_MODEL_LEN=1 VLLM_USE_FASTOKENS=1 vllm serve --model $MODEL_CKPT \
    --served-model-name nvidia/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-BF16 \
    --trust-remote-code \
    --max-num-seqs 512 \
    --max-model-len 1048576 \
    --max-num-batched-tokens 8192 \
    --enable-prefix-caching \
    --async-scheduling \
    --speculative_config.method dflash \
    --speculative_config.model $DFLASH_CKPT \
    --speculative_config.num_speculative_tokens 3 \
    --mamba-backend flashinfer \
    --mamba-ssm-cache-dtype float16 \
    --enable-mamba-cache-stochastic-rounding \
    --mamba-cache-philox-rounds 5

For more details on how to deploy and use the model — see the Quick Start Guide below!

License/Terms of Use

Governing Download Terms: Use of this model is governed by the OpenMDW-1.1 model license.

Benchmarks

We evaluated our model on the following benchmarks:

Agentic Coding Benchmarks

Additional harness-level coding-agent results for SWE-Bench Verified and Terminal-Bench 2.1 are summarized in the benchmark table above.

Deployment Geography: Global

Use Case

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.

Release Date

Hugging Face — 08/11/2026

Model Architecture

• Architecture Type: Mixture-of-Experts Hybrid (Mamba + Transformer)
• Network Architecture: Nemotron-3-Lightning + Multi-Token Prediction (MTP)
• Number of model parameters: 30B Total / 3B Active

Model Design

The model was pre-trained with over 20T tokens and supports up to 1M context length. The pre-training phase used an NVFP4 recipe. It utilizes the LatentMoE architecture, where tokens are projected into a smaller latent dimension for expert routing and computation, improving accuracy per byte. The model includes Multi-Token Prediction (MTP) layers, which predict multiple future tokens to provide richer training signals.

Training Methodology

Stage 1: Pre-Training

• 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.
• Software used for pre-training: Megatron-LM

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.
• Software used for reinforcement learning: NeMo RL, NeMo Gym

NVIDIA-Nemotron-3.5-Lightning-30B-A3B-BF16 is a result of the above work.

Input

• Input Type(s): Text
• Input Format(s): String
• Input Parameters: One-Dimensional (1D): Sequences
• Other Properties Related to Input: Maximum context length up to 1M tokens. Supported languages include English, Spanish, French, German, Italian, and Japanese.

Output

• Output Type(s): Text
• Output Format: String
• Output Parameters: One-Dimensional (1D): Sequences
• 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.

Software Integration

• Runtime Engine(s): PyTorch; llama.cpp (GGUF)
• Supported Hardware Microarchitecture Compatibility: NVIDIA Ampere - A100; NVIDIA Blackwell; NVIDIA Hopper - H100-80GB
• Preferred/Supported Operating System(s): Linux

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.

Model Version(s)

• GA (08/11/2026)

Quick Start Guide

All deployment snippets below assume:

export MODEL_CKPT=nvidia/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-BF16

And for DFlash:

export DFLASH_CKPT=nvidia/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-NVFP4-DFlash

The BF16 recipes below cover vLLM and llama.cpp. For TensorRT-LLM, SGLang, W4A16 (Blackwell / Hopper / Ampere from a single checkpoint), and DGX Spark (DSpark) recipes, see the NVFP4 card.

Speculative Decoding Strategies

Lightning 3.5 ships with external draft models for speculative decoding — pick the strategy that matches your deployment, or serve without speculative decoding for maximum-throughput batch workloads:

• DFlash: An external draft-model checkpoint tuned for low concurrency data centre deployments, minimizing per-request latency when batch sizes are small.
• DSpark: A draft-model configuration tuned specifically for DGX Spark (GB10) workflows. The DSpark recipe pairs with the NVFP4 (W4A16) checkpoint — see the NVFP4 card for the launch command.

The model's built-in Multi-Token Prediction (MTP) layers are used during training only; MTP-based serving is not recommended, and no MTP deployment configurations are provided.

The snippets in each backend section below show how to launch with the corresponding strategy.

vLLM

• Nightly container: vllm/vllm-openai:nightly-821717118fc26667dd474b9b0ab81d29259dfc5c

All vLLM snippets below serve on the default port 8000, matching the API Client examples.

1x H100

For high concurrency deployments, use the following:

VLLM_ALLOW_LONG_MAX_MODEL_LEN=1 VLLM_USE_FASTOKENS=1 vllm serve --model $MODEL_CKPT \
    --served-model-name nvidia/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-BF16 \
    --max-num-seqs 512 \
    --trust-remote-code \
    --max-model-len 262144 \
    --max-num-batched-tokens 16384 \
    --enable-prefix-caching \
    --async-scheduling \
    --mamba-backend flashinfer \
    --mamba-ssm-cache-dtype float16 \
    --enable-mamba-cache-stochastic-rounding \
    --mamba-cache-philox-rounds 5 \
    --mamba-cache-mode align

For low-concurrency deployments, you can use this following command, leveraging DFlash:

VLLM_ALLOW_LONG_MAX_MODEL_LEN=1 VLLM_USE_FASTOKENS=1 vllm serve --model $MODEL_CKPT \
    --served-model-name nvidia/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-BF16 \
    --trust-remote-code \
    --max-num-seqs 512 \
    --max-model-len 262144 \
    --max-num-batched-tokens 8192 \
    --enable-prefix-caching \
    --async-scheduling \
    --speculative_config.method dflash \
    --speculative_config.model $DFLASH_CKPT \
    --speculative_config.num_speculative_tokens 3 \
    --mamba-backend flashinfer \
    --mamba-ssm-cache-dtype float16 \
    --enable-mamba-cache-stochastic-rounding \
    --mamba-cache-philox-rounds 5

8x H100

For long-context, multi-GPU serving (TP8 with expert parallelism):

VLLM_ALLOW_LONG_MAX_MODEL_LEN=1 VLLM_USE_FASTOKENS=1 vllm serve --model $MODEL_CKPT \
    --served-model-name nvidia/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-BF16 \
    --moe-backend flashinfer_cutlass \
    --trust-remote-code \
    --max-num-batched-tokens 4096 \
    --mamba-backend flashinfer \
    --enable-prefix-caching \
    --mamba-cache-mode align \
    --max-model-len 1048576 \
    --enable-expert-parallel \
    --tensor-parallel-size 8

1x GB200

Specdec method - DFlash:

VLLM_ALLOW_LONG_MAX_MODEL_LEN=1 VLLM_USE_FASTOKENS=1 vllm serve --model $MODEL_CKPT \
    --served-model-name nvidia/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-BF16 \
    --max-num-seqs 64 \
    --max-model-len 1048576 \
    --max-num-batched-tokens 10240 \
    --no-enable-prefix-caching \
    --async-scheduling \
    --speculative_config.method dflash \
    --speculative_config.model $DFLASH_CKPT \
    --speculative_config.num_speculative_tokens 5 \
    --mamba-backend flashinfer

• Context Length: The snippets above serve the model's full 1M-token context window by default. 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.

Llama.cpp

Run Lightning 3.5 locally on consumer GPUs. The walkthrough below builds llama.cpp with DFlash support, converts the target and draft models to GGUF (with optional quantization), and serves an OpenAI-compatible endpoint on a single RTX 5090 — including a speed-bench harness to verify throughput.

git clone https://gitlab-master.nvidia.com/winai/llama.cpp.git -b lnigam/dflash-recurrent-rollback  # TODO: replace with public repo/branch before publish (checklist 8)

cd $LLAMA_CPP
 
cmake -S . -B build-cuda -G Ninja `
  -DCMAKE_BUILD_TYPE=Release `
  -DGGML_CUDA=ON
 
cmake --build build-cuda --config Release -j 18

# for target model
$PYTHON convert_hf_to_gguf.py $TARGET_HF `
  --outtype bf16 `
  --outfile "$OUT\nemotron-3.5-lightning-target-bf16.gguf"
 
# for draft model. Note that it requires the corresponding target model as an
# additional argument.
$PYTHON convert_hf_to_gguf.py $DFLASH_HF `
  --target-model-dir $TARGET_HF `
  --outtype bf16 `
  --outfile "$OUT\nemotron-3.5-dflash-bf16.gguf"
 
# one can also take NVFP4 models from HF and directly convert them into GGUF
# using the following command
$PYTHON convert_hf_to_gguf.py $MODEL_PATH `
  --fp8-as-q8 `
  --outfile model.gguf

"$LLAMA_CPP\build-cuda\bin\llama-quantize.exe" `
  "to\bf16\model.gguf" `
  "to\quantized\model.gguf" `
  Q4_K_M
 
"$LLAMA_CPP\build-cuda\bin\llama-quantize.exe" `
  "$OUT\nemotron-3.5-dflash-bf16.gguf" `
  "$OUT\nemotron-3.5-dflash-Q4_K_M.gguf" `
  Q4_K_M

"$LLAMA_CPP\build-cuda\bin\llama-server.exe" `
  -m "to\target\model.gguf" `
  --temp 0 -s 421 --top-k 1 `
  -np 1 `
  -c 40960 `
  --port $PORT `
  -ngl 99 `
  -fa on `
  --jinja `
  --no-webui `
  --fit off

"$LLAMA_CPP\build-cuda\bin\llama-server.exe" `
  -m "to\target\model.gguf" `
  -md "to\draft\model.gguf" `
  --spec-type draft-dflash `
  --spec-draft-n-max 7 `
  --temp 0 -s 421 --top-k 1 `
  -np 1 `
  -c 40960 `
  --port $PORT `
  -ngl 99 `
  -fa on `
  --jinja `
  --no-webui `
  --fit off

# run this once
pip install -r "$LLAMA_CPP\tools\server\bench\speed-bench\requirements.txt"
 
# do a hf login, if the following command doesn't work.
$PYTHON "$LLAMA_CPP\tools\server\bench\speed-bench\speed_bench.py" `
  --url http://127.0.0.1:$PORT `
  --bench qualitative `
  --category all `
  --osl 128 `
  --concurrency 1 `
  --output "$OUT\result.json"

API Client

The examples below use the OpenAI-compatible client and work with either serving backend above. Both backends serve on port 8000 (vLLM by default; llama.cpp via $PORT=8000), so the base_url works as-is. Recommended sampling settings are Temperature 1.0 and Top_P 0.95.

The vLLM snippets above register the model as nvidia/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-BF16 via --served-model-name. For llama.cpp — or if you change that flag — 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), thinking disabled for direct answers, and a runtime thinking budget.

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)

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)

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)

Tool Calling

For vLLM, add the following to any serve command above:

    --enable-auto-tool-choice \
    --tool-call-parser qwen3_coder \
    --reasoning-parser nemotron_v3

NOTE: For coding agents, add extra_body={"chat_template_kwargs": {"force_nonempty_content": True}} to the API call, as shown below.

tools = [{
    "type": "function",
    "function": {
        "name": "get_weather",
        "description": "Get the current weather for a city",
        "parameters": {
            "type": "object",
            "properties": {"city": {"type": "string"}},
            "required": ["city"],
        },
    },
}]

response = client.chat.completions.create(
    model=MODEL,
    messages=[{"role": "user", "content": "What's the weather in Santa Clara?"}],
    tools=tools,
    max_tokens=16000,
    temperature=1.0,
    top_p=0.95,
    extra_body={"chat_template_kwargs": {"force_nonempty_content": True}},
)
print(response.choices[0].message.tool_calls)

Training, Testing, and Evaluation Datasets

Training

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.

Testing Datasets:

Data Collection Method by dataset

• Hybrid: Automated, Manually-Collected, Synthetic Labeling Method by dataset
  
• Hybrid: Automated, Manually-Labeled, Synthetic Properties: This corpus comprises a mix of high-quality standard benchmarks and test suites for modern agentic AI. These benchmarks test model capabilities on tasks such as tool-calling and instruction following.

Evaluation Datasets:

Data Collection Method by dataset

• Hybrid: Automated, Manually-Collected, Synthetic Labeling Method by dataset
  
• Hybrid: Automated, Manually-Labeled, Synthetic Properties: This corpus comprises a mix of high-quality standard benchmarks and test suites for modern agentic AI. These benchmarks test model capabilities on tasks such as tool-calling and instruction following.

Inference

• Acceleration Engine: PyTorch
• Test Hardware:
  • NVIDIA Hopper
    • 1-8x H100
    • 1-8x H200
  • NVIDIA Blackwell
    • GB200
    • GeForce RTX 5090

Ethical Considerations

NVIDIA believes Trustworthy AI is a shared responsibility and we have established policies and practices to enable development for a wide array of AI applications. Developers should work with their internal model team to ensure this model meets requirements for the relevant industry and use case and addresses unforeseen product misuse.

We advise against circumvention of any provided safety guardrails contained in the Model without a substantially similar guardrail appropriate for your use case. For more details, see the Safety and Security and Explainability tabs on this NGC model page.

For more detailed information on ethical considerations for this model, please see the Bias and Privacy tabs on this NGC model page.

Please report model quality, risk, security vulnerabilities or NVIDIA AI Concerns here.

Get Help

Getting started with the NIM

Deploying and integrating the NIM is straightforward thanks to our industry standard APIs. Visit the NVIDIA NIM documentation for release documentation, deployment guides and more.

NVIDIA Developer Community Forum

Get access to community knowledge base articles and support cases (https://forums.developer.nvidia.com/)