Nvidia Nemotron 3.5 Lightning

Model Summary

Model Overview

Model Developer: NVIDIA Corporation

Model Dates: December 2025 - May 2026

Data Freshness:

• The pre-training information has a cutoff day of September 2025.
• The post-training information has a cutoff day of May 2026.

What is Nemotron?

NVIDIA Nemotron™ is simply a family of unfastened models pinch unfastened weights, training data, and recipes, delivering starring ratio and accuracy for building specialized AI agents.

Description

NVIDIA-Nemotron-3.5-Lightning-30B-A3B-NVFP4 is simply a ample connection exemplary (LLM) trained by NVIDIA.

The exemplary employs a hybrid Mixture-of-Experts architecture, utilizing interleaved Mamba-2 and MoE layers, on pinch prime Attention layers. The Lightning 3.5 exemplary is released alongside a number of speculative decoding methods for faster matter generation. The exemplary has 3B progressive parameters and 30B parameters successful total.

This exemplary is fresh for commercialized use.

Quick Start

To get quickly started connected DGX Spark (GB10) you tin usage the pursuing command.

Grab the model:

export MODEL_CKPT=nvidia/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-NVFP4 export DSPARK_CKPT=nvidia/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-NVFP4-DSpark

Run it pinch vLLM — this look uses DSpark speculative decoding, tuned for DGX Spark. (vLLM Nightly: vllm/vllm-openai:v0.27.1)

vllm service --model $MODEL_CKPT \ --moe-backend marlin \ --kv-cache-dtype fp8 \ --enable-prefix-caching \ --speculative_config.num_speculative_tokens 3 \ --mamba-backend flashinfer \ --mamba-cache-mode align \ --reasoning-parser nemotron_v3 \ --speculative_config.model $DSPARK_CKPT \ --tool-call-parser qwen3_coder \ --enable-auto-tool-choice

For much specifications connected really to deploy and usage the exemplary — spot the Quick Start Guide below!

License/Terms of Use

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

Benchmarks

Reasoning Benchmark Evaluations

We evaluated our exemplary connected the pursuing benchmarks:

Accuracy numbers measured by NVIDIA nether a accordant harness (NeMo Gym / Nemo Evaluator SDK); they whitethorn disagree from vendors' self-reported numbers.

For reproducibility, the information recipes, installation instructions, and commands for NVIDIA Nemotron 3.5 Lightning were collected and published successful NeMo Gym. The reported results screen the merchandise information suite, including knowledge and reasoning, instruction following, coding, agentic, tool-use, and long-context. Most evaluations usage NeMo Gym-native harnesses while a mini subset, including SWE-Bench and Terminal-Bench, utilized NeMo Evaluator natively. The published recipes specify the benchmark-specific containers, prompts, conclusion parameters, parser configurations, and scoring settings utilized to nutrient the results.

These numbers were measured pinch and use to the charismatic NVFP4 checkpoint

Agentic Coding Benchmarks

Additional harness-level coding-agent results for SWE-Bench Verified and Terminal-Bench 2.1 are shown below.

Deployment Geography: Global

Use Case

NVIDIA-Nemotron-3.5-Lightning-30B-A3B-NVFP4 is simply a wide intent reasoning and chat exemplary intended to beryllium utilized successful English and coding languages. Other non-English languages (Spanish, French, German, Italian, Japanese) are besides supported. Intended for developers designing AI Agent systems, chatbots, RAG systems, and different AI-powered applications. Also suitable for emblematic instruction-following tasks.

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 exemplary parameters: 30B Total / 3B Active

Model Design

The exemplary was pre-trained pinch complete 20T tokens and supports up to 1M discourse length. The pre-training shape utilized an NVFP4 recipe. The exemplary includes Multi-Token Prediction (MTP) layers, which foretell aggregate early tokens to supply richer training signals.

Training Methodology

Stage 1: Pre-Training

• NVIDIA-Nemotron-3.5-Lightning-30B-A3B-NVFP4 exemplary was pre-trained utilizing an NVFP4 look pinch crawled and synthetic code, math, science, and wide knowledge data.
• Software utilized for pre-training: Megatron-LM

Stage 2: Continued Pre-Training for Multi-Token Prediction (MTP)

• The exemplary underwent a continued pre-training shape to train its Multi-Token Prediction (MTP) layers. In this stage, MTP heads study to foretell aggregate early tokens, providing richer training signals to the guidelines model. This shape aligns the MTP layers pinch the guidelines model's distribution.

Stage 3: Supervised Fine-Tuning

• The exemplary was further fine-tuned connected synthetic code, math, science, instrumentality calling, instruction following, system outputs, and wide knowledge data. This shape incorporated information designed to support long-range retrieval and multi-document aggregation.

Stage 4: Reinforcement Learning

• The exemplary underwent multi-environment reinforcement learning utilizing GRPO (Group Relative Policy Optimization) crossed math, code, science, instruction following, multi-step instrumentality use, multi-turn conversations, and system output environments. It utilized an asynchronous RL architecture that decouples training from conclusion and leverages MTP to accelerate rollout generation.
• Software utilized for reinforcement learning: NeMo RL, NeMo Gym

Stage 5: Post-training Quantization (PTQ)

• We performed post-training quantization (PTQ) pinch Nvidia Model Optimizer utilizing the pursuing recipe: Four Over Six NVFP4 (a version of fixed MSE calibration) W4A16 connected routed and shared experts, FP8 per-tensor move scales connected mamba in_proj/out_proj and KV cache. We utilized a subset of the Nemotron Ultra validation group for calibration pinch 1000 samples astatine 32k token length.

NVIDIA-Nemotron-3.5-Lightning-30B-A3B-NVFP4 is simply a consequence of the supra work.

Input

• Input Type(s): Text
• Input Format(s): String
• Input Parameters: One-Dimensional (1D): Sequences
• Other Properties Related to Input: Maximum discourse magnitude up to 1M tokens. Supported languages see 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 discourse magnitude up to 1M tokens

Our AI models are designed and/or optimized to tally connected NVIDIA GPU-accelerated systems. By leveraging NVIDIA's hardware (e.g. GPU cores) and package frameworks (e.g., CUDA libraries), the exemplary achieves faster training and conclusion times compared to CPU-only solutions.

Software Integration

• Runtime Engine(s): PyTorch
• Supported Hardware Microarchitecture Compatibility: NVIDIA Blackwell; NVIDIA Hopper (NVFP4 / W4A16); NVIDIA Ampere (W4A16)
• Preferred/Supported Operating System(s): Linux

The integration of instauration and fine-tuned models into AI systems requires further testing utilizing use-case-specific information to guarantee safe and effective deployment. Following the V-model methodology, iterative testing and validation astatine some portion and strategy levels are basal to mitigate risks, meet method and functional requirements, and guarantee compliance pinch information and ethical standards earlier deployment.

Model Version(s)

• GA (08/11/2026)

Quick Start Guide

All deployment snippets beneath assume:

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

And for DSpark:

export DSPARK_CKPT=nvidia/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-NVFP4-DSpark

Speculative Decoding Strategies

Lightning 3.5 ships pinch 2 outer draught models for speculative decoding arsenic good arsenic MTP (Multi-Token Prediction). While we presently urge DSpark for each cases - your usecase whitethorn align pinch DFlash and MTP:

• DSpark: A semi-autoregressive speculative-decoding drafter that proposes a full artifact of campaigner tokens successful a azygous guardant walk from a parallel backbone. This is recommended for DGX Spark, arsenic good arsenic low-concurrency information centre deployments.
• DFlash: A speculative-decoding drafter that uses a lightweight block-diffusion exemplary to make an full draught artifact successful 1 guardant pass.
• MTP: A modeling method that trains the web to foretell respective early tokens astatine each position alternatively of only the adjacent one.

vLLM

For much indepth instructions connected really to deploy done vLLM, caput here

• vLLM Nightly: vllm/vllm-openai:v0.27.1

1x DGX Spark (GB10)

Specdec method - DSpark:

vllm service --model $MODEL_CKPT \ --moe-backend marlin \ --kv-cache-dtype fp8 \ --max-model-len 1048576 \ --enable-prefix-caching \ --speculative_config.num_speculative_tokens 3 \ --mamba-backend flashinfer \ --mamba-cache-mode align \ --reasoning-parser nemotron_v3 \ --speculative_config.method dspark \ --tool-call-parser qwen3_coder \ --enable-auto-tool-choice

1x H100

For max throughput deployments, usage the pursuing configuration, nary speculative decoding strategy is champion for this serving configuration, and owed to representation constraints the Mamba cache dtype is group arsenic FP16:

vllm service --model $MODEL_CKPT \ --max-num-seqs 256 \ --max-num-batched-tokens 16384 \ --enable-prefix-caching \ --async-scheduling \ --mamba-backend flashinfer \ --moe-backend humming \ --linear-backend humming \ --mamba-ssu-algorithm horizontal \ --mamba-cache-mode align \ --mamba-ssm-cache-dtype float16 \ --enable-mamba-cache-stochastic-rounding \ --mamba-cache-philox-rounds 5 \ --reasoning-parser nemotron_v3 \ --tool-call-parser qwen3_coder \ --enable-auto-tool-choice

For interactive usage scenarios (achieving 40+ TPS/User) usage a little concurrency (<=128) pinch DSpark:

vllm service --model $MODEL_CKPT \ --max-num-seqs 128 \ --enable-prefix-caching \ --async-scheduling \ --speculative_config.model $DSPARK_CKPT \ --speculative_config.num_speculative_tokens 3 \ --mamba-ssu-algorithm horizontal \ --mamba-backend flashinfer \ --mamba-ssm-cache-dtype float16 \ --enable-mamba-cache-stochastic-rounding \ --mamba-cache-philox-rounds 5 \ --reasoning-parser nemotron_v3 \ --tool-call-parser qwen3_coder \ --enable-auto-tool-choice

8x H100

For long-context, multi-GPU serving (TP8 pinch master parallelism):

vllm service --model $MODEL_CKPT \ --mamba-backend flashinfer \ --async-scheduling \ --enable-prefix-caching \ --mamba-cache-mode align \ --enable-expert-parallel \ --tensor-parallel-size 8 \ --reasoning-parser nemotron_v3 \ --tool-call-parser qwen3_coder \ --enable-auto-tool-choice

1x GB200

vllm service --model $MODEL_CKPT \ --max-num-batched-tokens 10240 \ --no-enable-prefix-caching \ --async-scheduling \ --speculative_config.model $DSPARK_CKPT \ --speculative_config.num_speculative_tokens 5 \ --mamba-backend flashinfer \ --reasoning-parser nemotron_v3 \ --tool-call-parser qwen3_coder \ --enable-auto-tool-choice

W4A16 — Ampere

The aforesaid checkpoint besides serves via W4A16 kernels, extending sum to Ampere-class GPUs:

vllm service --model nvidia/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-NVFP4 \ --moe-backend humming \ --linear-backend humming \ --max-num-seqs 256 \ --max-num-batched-tokens 32768 \ --enable-prefix-caching \ --async-scheduling \ --quantization modelopt_fp4 \ --mamba-backend flashinfer \ --mamba-cache-mode align \ --mamba-ssu-algorithm elemental \ --reasoning-parser nemotron_v3 \ --tool-call-parser qwen3_coder \ --enable-auto-tool-choice

• Context Length: The H100 and GB200 snippets supra service the model's afloat 1M-token discourse model by default. If you're memory-constrained — aliases want much KV-cache headroom astatine precocious concurrency — little --max-model-len to lucifer your workload.

TensorRT-LLM

For much indepth instructions connected really to deploy done TensorRT-LLM, caput here

Container: nvcr.io/nvidia/tensorrt-llm/release:1.3.0rc24

1x H100

cat > nemotron-35-lightning-nvfp4-mtp.yaml << EOF kv_cache_config: dtype: fp8 enable_block_reuse: false mamba_state_config: periodic_snapshot_interval: 8192 free_gpu_memory_fraction: 0.8 mamba_ssm_cache_dtype: float16 mamba_ssm_stochastic_rounding: true mamba_ssm_philox_rounds: 5 moe_config: backend: MARLIN nvfp4_gemm_config: allowed_backends: [marlin, cutlass, cublaslt, cuda_core] cuda_graph_config: enable_padding: true max_batch_size: 8 speculative_config: decoding_type: MTP max_draft_len: 3 allow_advanced_sampling: true enable_chunked_prefill: true num_postprocess_workers: 4 print_iter_log: true stream_interval: 10 disable_overlap_scheduler: false EOF trtllm-serve \ $MODEL_CKPT \ --max_batch_size 8 \ --max_num_tokens 8192 \ --reasoning_parser nemotron-v3 \ --tool_parser qwen3_coder \ --config nemotron-35-lightning-nvfp4-mtp.yaml

• Context length: The bid supra serves the model's afloat 1M-token discourse model by default. If you're memory-constrained — aliases want much KV-cache headroom astatine higher concurrency — little --max_seq_len to lucifer your workload.

SGLang

For much indepth instructions connected really to deploy done SGLang, caput here

• Container: lmsysorg/sglang:dev-nemotron3-5-lighting

1x H100

sglang service \ --model-path nvidia/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-NVFP4 \ --max-running-requests 256 \ --trust-remote-code \ --chunked-prefill-size 32768 \ --mem-fraction-static 0.9 \ --speculative-algorithm EAGLE \ --speculative-draft-model-path nvidia/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-NVFP4 \ --speculative-num-steps 3 \ --speculative-eagle-topk 1 \ --speculative-num-draft-tokens 4 \ --mamba-backend flashinfer \ --mamba-radix-cache-strategy extra_buffer \ --reasoning-parser nemotron_3 \ --tool-call-parser qwen3_coder

• Context length: The bid supra serves the model's afloat 1M-token discourse model by default. If you're memory-constrained — aliases want much KV-cache headroom astatine higher concurrency — group --context-length to a smaller value.

API Client

The examples beneath usage the OpenAI-compatible customer and activity pinch immoderate of the serving backends above. All backends service connected larboard 8000 (vLLM and TRT-LLM by default; SGLang via $PORT=8000), truthful the base_url useful as-is. Recommended sampling settings are Temperature 1.0 and Top_P 0.95.

The vLLM snippets supra registry the exemplary arsenic nvidia/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-NVFP4 via --served-model-name. For the different backends — aliases if you alteration that emblem — transcript 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-NVFP4"

Lightning 3.5 exposes reasoning power done chat-template kwargs: reasoning enabled (the default), and reasoning abnormal for nonstop 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 astir 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 superior 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 successful 2 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

The TRT-LLM snippet supra already launches pinch the required parsers (--reasoning_parser nemotron-v3 --tool_parser qwen3_coder). For vLLM, adhd the pursuing to immoderate service bid above:

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

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

tools = [{ "type": "function", "function": { "name": "get_weather", "description": "Get the existent upwind 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 upwind successful 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

Data Modality: Text Training Data Size: More than 20 Trillion Tokens Dataset partition: Training [100%], testing [0%], validation [0%] Time play for training information collection: 2013 to December 2025 Time play for testing information collection: 2013 to December 2025 Time play for validation information 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-NVFP4 is pre-trained connected a ample corpus of high-quality curated and synthetically-generated data. It is trained successful the English language, arsenic good arsenic 19 different spoken languages and 43 programming languages. Our sources screen a assortment of archive types specified as: webpages, dialogue, articles, and different written materials. The corpus spans domains including legal, math, science, finance, and more. We besides see a mini information of question-answering, and alignment style information to amended exemplary accuracy. The exemplary was pre-trained for much than 20 trillion tokens.

The post-training corpus for NVIDIA-Nemotron-3.5-Lightning-30B-A3B-NVFP4 consists of high-quality curated and synthetically-generated data. Primary languages utilized for post-training see English, French, German, Italian, Japanese, Spanish, and Chinese.

These datasets, specified arsenic FinePDFs, EssentialWeb, HotpotQA, SQuAD, and HelpSteer3, do not collectively aliases exhaustively correspond each demographic groups (and proportionally therein). For instance, these datasets do not incorporate definitive mentions of demographic classes specified arsenic age, gender, aliases ethnicity successful 64-99% of samples, depending connected the source. In the subset wherever specified position are present, document-based datasets (FinePDFs and EssentialWeb) incorporate representational skews, specified arsenic references to "male" outnumbering those to "female", and mentions of "White" arsenic the astir predominant among taste identifiers (comprising 43-44% of ethnicity mentions). To mitigate these imbalances, we urge considering information techniques specified arsenic bias audits, fine-tuning pinch demographically balanced datasets, and mitigation strategies for illustration counterfactual information augmentation to align pinch the desired exemplary behavior. This information utilized a 3,000-sample subset per dataset, identified arsenic the optimal period for maximizing embedder accuracy.

During post-training, we make synthetic information by distilling trajectories, solutions, and translations from beardown coach models and supplier systems, often grounded successful existent tasks aliases documents and aggressively filtered for quality. For math, code, and science, we commencement from curated problem sets and usage unfastened root permissive models specified arsenic GPT-OSS-120B to nutrient step-by-step reasoning traces, campaigner solutions, best-of-n action traces, and verified CUDA kernels. For long-context and science, we build synthetic QA and reasoning information by retrieving passages from agelong documents, generating MCQ/OpenQA questions and answers, and paraphrasing them into aggregate prompt/response formats to guarantee diversity. Across each pipelines we stack automated verification—compilers, numerical checks, connection identification—to guarantee our information is precocious quality.

For each domains, we use a unified information filtering pipeline to guarantee that only high-quality, license-compliant, and verifiable samples are utilized for post-training. We first discard malformed examples utilizing structural checks (e.g., missing instrumentality definitions erstwhile instrumentality calls are present). We past aggressively select reasoning traces exhibiting pathological repetition, specified arsenic repeated n-grams wrong a sliding model aliases crossed the full trajectory, which we recovered to beryllium a beardown parameter of malformed aliases low-quality reasoning. Finally, based connected soul audits of synthetically generated datasets, we observed that immoderate coach models occasionally nutrient reasoning traces and last responses that implicitly align pinch circumstantial governmental entities aliases beforehand nationalistic narratives. To mitigate this, we use targeted keyword- and regex-based filters and region each trajectories matching specified behavior.

Alongside the model, we merchandise our last pre-training and post-training data, arsenic outlined successful this section. For easiness of analysis, location is simply a sample group that is ungated. For each remaining code, mathematics and multilingual data, gating and support is required, and the dataset is permissively licensed for exemplary training purposes.

For Detailed Dataset Information: Click here!

Base Pre-Training Corpus (Nemotron 3 Foundation)

The instauration of the exemplary is trained connected the Nemotron 3 corpus, comprising the pursuing datasets from the Nemotron Pretraining Datasets collection:

Public Datasets

Crawled and Scraped from Online Sources by NVIDIA

The English Common Crawl information was downloaded from the Common Crawl Foundation (see their FAQ for specifications connected their crawling) and includes the snapshots CC-MAIN-2013-20 done CC-MAIN-2025-13. The information was subsequently deduplicated and filtered successful various ways described successful the Nemotron-CC paper. Additionally, we extracted information for 15 languages from the pursuing 3 Common Crawl snapshots: CC-MAIN-2024-51, CC-MAIN-2025-08, CC-MAIN-2025-18. The 15 languages included were Arabic, Chinese, Danish, Dutch, French, German, Italian, Japanese, Korean, Polish, Portuguese, Russian, Spanish, Swedish, and Thai. As we did not person reliable multilingual model-based value classifiers available, we applied conscionable heuristic filtering instead—similar to what we did for little value English information successful the Nemotron-CC pipeline, but selectively removing immoderate filters for immoderate languages that did not activity well. Deduplication was done successful the aforesaid measurement arsenic for Nemotron-CC.

The GitHub Crawl was collected utilizing the GitHub REST API and the Amazon S3 API. Each crawl was operated successful accordance pinch the complaint limits group by its respective source, either GitHub aliases S3. We cod earthy root codification and subsequently region immoderate having a licence which does not beryllium successful our permissive-license set.

Private Non-publicly Accessible Datasets of Third Parties

Private Non-publicly Accessible Datasets by NVIDIA

NVIDIA-Sourced Synthetic Datasets (Pre-Training)

NVIDIA-Sourced Synthetic Datasets (Post-Training)

Language Distribution successful Post-Training

For our post-training recipe, we focused connected the pursuing languages successful summation to English: French, German, Italian, Japanese, Spanish, and Chinese. Those languages were represented successful the shape of multilingual reasoning and translator tasks.

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 operation of high-quality modular benchmarks and trial suites for modern agentic AI. These benchmarks trial exemplary capabilities connected tasks specified arsenic 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 operation of high-quality modular benchmarks and trial suites for modern agentic AI. These benchmarks trial exemplary capabilities connected tasks specified arsenic tool-calling and instruction following.

Inference

• Acceleration Engine: PyTorch
• Test Hardware:
  • NVIDIA Hopper
    • 1-8x H100
    • 1-8x H200
  • NVIDIA Blackwell
    • GB200
    • DGX Spark (GB10)
    • GeForce RTX 5090

Ethical Considerations

NVIDIA believes Trustworthy AI is simply a shared work and we person established policies and practices to alteration improvement for a wide array of AI applications. Developers should activity pinch their soul exemplary squad to guarantee this exemplary meets requirements for the applicable manufacture and usage lawsuit and addresses unforeseen merchandise misuse.

We counsel against circumvention of immoderate provided information guardrails contained successful the Model without a substantially akin guardrail due for your usage case. For much details: Safety and Explainability Subcards.

For much elaborate accusation connected ethical considerations for this model, please spot the Model Card++ Bias, and Privacy Subcards.

Please study exemplary quality, risk, information vulnerabilities aliases NVIDIA AI Concerns here.