NVIDIA-Nemotron-3.5-Bolt-74B-A9B-24GB-LiLiCorr

Description

NVIDIA-Nemotron-3.5-Bolt-74B-A9B-24GB-LiLiCorr is the LiLiCorr speculative-decoding checkpoint for the NVIDIA-Nemotron-3.5-Bolt-74B-A9B-24GB. The compatible target is a hybrid LatentMoE vision-language model for reasoning, chat, and agent workflows. For more information, see the NVIDIA-Nemotron-3.5-Bolt-74B-A9B-24GB target card.

This checkpoint supports LiLiCorr-assisted serving of the compatible NVIDIA-Nemotron-3.5-Bolt-74B-A9B-24GB target.

This checkpoint is not a standalone language model. It proposes draft token candidates that the compatible 24GB target verifies; the target produces the user-visible text.

This model is ready for commercial or non-commercial use.

License/Terms of Use

GOVERNING DOWNLOAD TERMS: Use of this model is governed by the OpenMDW-1.1 model license.

Deployment Geography

Global

Use Case

Developers can pair this draft with NVIDIA-Nemotron-3.5-Bolt-74B-A9B-24GB for text generation. The target supports reasoning, chat, retrieval-augmented generation, visual understanding, and agent workflows. The draft accelerates generation through speculative decoding.

Release Date

Hugging Face 10/15/2026 via https://huggingface.co/nvidia/NVIDIA-Nemotron-3.5-Bolt-74B-A9B-24GB-LiLiCorr

References

Model Architecture

Architecture Type: Dense GQA (Dense MLP + GQA Attention)

Network Architecture: Five dense feed-forward layers use grouped-query attention (GQA). Draft attention is bidirectional within the draft block. All five layers use a 1024-token per-head sliding window and attention sinks, with a grouped dynamic convolution around each attention and MLP sublayer. A separate two-layer LiLiCorr head correlates the drafter's top-8 candidates at each draft slot.

Number of Model Parameters: 1.113B draft parameters, excluding shared target embeddings and the LM head. The two-layer LiLiCorr head accounts for 37M of them.

Target Model Architecture: Hybrid LatentMoE vision-language model with 74B total parameters and 9B active parameters.

Input

The inputs below describe the target-model serving interface. LiLiCorr drafts tokens using representations from the target model.

Input Type(s): Text Input Format(s): Text: String Input Parameters: Text: One-Dimensional (1D) Other Properties Related to Input: The target model supports up to 256K tokens of context. Supported languages include English, French, Spanish, Italian, German, Japanese, Hindi, Korean, Brazilian Portuguese, and Chinese.

Output

Output Type(s): Text
Output Format: String
Output Parameters: One-Dimensional (1D): Sequences
Other Properties Related to Output: Maximum context length up to 256K tokens. Supports JSON output format, reasoning output with chain-of-thought, and tool calling.

Other Properties Related to Output: Outputs can include responses, reasoning traces, tool-use content, and structured data. The chat template and application configuration determine the output form.

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

Supported Runtime Engine(s):

vLLM

Supported Hardware Microarchitecture Compatibility:

NVIDIA Blackwell, SM 120 only, including DGX Spark (GB10)

Preferred 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)

The model is a NVIDIA-Nemotron-3.5-Bolt-74B-A9B-24GB-LiLiCorr speculative decoding version quantized with NVIDIA Model Optimizer @ b5f10f8 using NVFP4 weights.

Training and Evaluation Datasets

Training Dataset

For target training details, see the NVIDIA-Nemotron-3.5-Bolt-74B-A9B-24GB target card.

Link: Nemotron-Post-Training-Dataset-v2, Nemotron-RL-Super-Training-Blends, Nemotron-Post-Training-v3, Nemotron-Instruction-Following-Chat-v1, and Nemotron-Science-v1. Only dataset prompts and conversation context were used for data synthesis. The original responses were not training targets. Every training response was generated by Nemotron-3.5-Bolt at temperature 1.0. The draft was first trained on responses from a Nemotron-3.5-Bolt post-training checkpoint, then fine-tuned on responses from the released checkpoint.

Data Modality: Text

Training Data Size: 1 Billion to 10 Trillion Tokens

Data Collection Method by dataset: Hybrid: Automated, manually-collected, synthetic.

Labeling Method by dataset: Hybrid: Automated, manually-labelled, synthetic.

Properties: This model was trained exclusively on a post-training corpus includes synthetic data for reasoning, code, science, tool use, instruction following, structured outputs, and multilingual tasks.

For more information about the target model's training data, see the Public Summary of Training Content.

Evaluation Dataset

The model is evaluated on SPEED-Bench for speculative decoding.

Data Collection Method by dataset: Hybrid: Automated, manually-collected, synthetic.

Labeling Method by dataset: Hybrid: Automated, manually-labelled, synthetic.

Properties: This corpus comprises a mix of high-quality standard benchmarks and test suites for LLMs. These benchmarks prompt the underlying model on a diverse set of tasks such as coding, math, writing, translation, etc., and the acceptance rates achieved from speculative decoding over the responses are measured in order to evaluate the quality of the speculation model.

Inference

Acceleration Engine: vLLM

Test Hardware: NVIDIA DGX Spark (GB10)

LiLiCorr Speculative Decoding

This release provides the LiLiCorr draft for the compatible Nemotron-3.5-Bolt-74B-A9B-24GB target. LiLiCorr proposes token blocks that the target verifies during generation.

This model supports a speculative decoding path with LiLiCorr for DGX Spark and low-concurrency data center workflows. This release is the LiLiCorr speculative decoding checkpoint intended to be paired with the NVIDIA-Nemotron-3.5-Bolt-74B-A9B-24GB target model. LiLiCorr is designed to improve accepted length while preserving the latency benefits of block drafting on compact Blackwell systems such as DGX Spark.

Usage

For serving with vLLM, see the compatible Nemotron-3.5-Bolt-74B-A9B-24GB target card.

Evaluation

Acceptance length on SPEED-Bench with draft length 7:

Category	SPEED-Bench Acceptance Length
coding	5.26
humanities	3.54
math	4.59
multilingual	5.06
qa	3.91
rag	5.13
reasoning	4.47
roleplay	3.45
stem	3.73
summarization	4.80
writing	3.41
Overall Average	4.30

Baseline: Nemotron-3.5-Bolt-74B-A9B-24GB Benchmarked with temperature=1.0, top_p=0.95.

These are checkpoint-level acceptance results and are not attributed to a specific serving runtime.

Model Limitations

This checkpoint is not a standalone language model and cannot produce user-visible output without the compatible 24GB target. The draft checkpoint provides no independent safety controls; retain the target model's safeguards and application guardrails.

The target model may generate inaccurate, incomplete, or otherwise undesirable responses, even when prompts are benign. It may reflect biases or content artifacts present in its training data, and output quality can vary by domain, language, reasoning depth, and context length. Developers should add application-specific safeguards and evaluate the system in the intended deployment environment before production use.

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.

For more detailed information on ethical considerations for this model, see the Model Card++ Bias, Explainability, Safety & Security, and Privacy Subcards.

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

SUBCARDS:

Explainability

Field	Response
Intended Task/Domain	Text generation from text, image, and video inputs; reasoning; tool use; and agentic workflows accelerated with speculative decoding
Model Type	Hybrid LatentMoE vision-language model used with a LiLiCorr draft checkpoint
Intended Users	Developers who use the compatible 24GB target for reasoning, chat, retrieval, visual understanding, and agent workflows
Output	Draft token candidates; the paired target model produces user-visible text
Describe how the model works	The target uses Mamba-2, MoE, and attention layers. LiLiCorr proposes token blocks. The target verifies each block.
Name the adversely impacted groups this has been tested to deliver comparable outcomes regardless of	Not Applicable
Technical Limitations & Mitigation	With correct target-model verification, speculative decoding preserves the target model's output distribution. Acceptance length and latency depend on the workload, sampling settings, hardware, and serving configuration.
Verified to have met prescribed quality standards?	Acceptance was evaluated on SPEED-Bench. Deployment quality requires separate evaluation.
Performance Metrics	Accuracy, throughput, latency, and speculative-decoding acceptance length
Potential Known Risk	The target model was trained on data that contains toxic language and societal biases originally crawled from the internet. Therefore, the model may amplify those biases and return toxic responses especially when prompted with toxic prompts. Therefore, before deploying any applications of this model, developers should perform safety testing and tuning tailored to their specific applications of the model. Developers should validate task accuracy, latency, and safety in their own deployment environment and add application-specific safeguards before production use.
Licensing	Governing Terms: Use of this model is governed by the OpenMDW-1.1 model license

Bias

Field	Response
Participation considerations from adversely impacted groups protected classes in model design and testing	None
Measures taken to mitigate against unwanted bias	None
Bias Metric	None

Safety & Security

Field	Response
Model Application(s)	Chat, instruction following, RAG, reasoning, visual understanding, and agentic AI workflows
Describe life critical application (if present)	Not Applicable
Use Case Restrictions	Use must comply with the OpenMDW-1.1 model license and applicable laws and regulations
Model and Dataset Restrictions	The Principle of least privilege (PoLP) is applied limiting access for dataset generation and model development. Restrictions enforce dataset access during training, and dataset license constraints adhered to.

Privacy

Field	Response
Generatable or Reverse engineerable personal data?	No
Personal data used to create this model?	No
Was consent obtained for any personal data used?	Not Applicable
How often is dataset reviewed?	Before Release
Was data from user interactions with the AI model (e.g. user input and prompts) used to train the model?	No
Is there provenance for all datasets used in training?	Yes
Does data labeling (annotation, metadata) comply with privacy laws?	Yes
Is data compliant with data subject requests for data correction or removal, if such a request was made?	Not Applicable
Applicable NVIDIA Privacy Policy	https://www.nvidia.com/en-us/about-nvidia/privacy-policy/