Model Overview

Description:

The NVIDIA Muse-Spark-Open-816B-A42B-DFlash2 model is the DFlash 2 draft checkpoint for Muse-Spark-Open-816B-A42B, a multimodal mixture-of-experts causal language model that produces text and supports long-context, tool-enabled workflows. The NVIDIA Muse-Spark-Open-816B-A42B-DFlash2 model incorporates DFlash 2 speculative decoding with Model Optimizer. DFlash 2 adds a two-tap in-block convolution around every sublayer and a pairwise path selector that walks one coherent path through each position’s top-k candidates.

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

Third-Party Community Consideration:

The target model, Muse-Spark-Open-816B-A42B, is not owned or developed by NVIDIA. It has been developed and built to a third-party's requirements; see link to Non-NVIDIA Muse-Spark-Open-816B-A42B Model Card. This DFlash 2 draft model was trained by NVIDIA for use with that target model.

License/Terms of Use:

GOVERNING TERMS: Use of the model/checkpoints is governed by the NVIDIA Open Model Agreement.
Additional Information: Meta’s Muse Model License. Built with Muse Spark Open.

Deployment Geography:

Global

Use Case:

Muse-Spark-Open-816B-A42B is intended for evaluation and development of long-context multimodal and agentic text-generation workflows, including tool-assisted coding. This DFlash2 release is intended for DFlash 2–assisted serving of Muse-Spark-Open-816B-A42B rather than as a standalone target model checkpoint.

Release Date:

Huggingface: 10/16/2026 via https://huggingface.co/nvidia/Muse-Spark-Open-816B-A42B-DFlash2

References

NVIDIA Model Optimizer: https://github.com/NVIDIA/Model-Optimizer DFlash: Block Diffusion for Flash Speculative Decoding

Model Architecture:

Architecture Type: Transformer
Network Architecture: 5 dense decoder layers with non-causal sliding-window attention (window 2048)
Total Parameters: 4.38B; the token embedding and output head are shared with the target and not included
Active Parameters: 4.38B (dense, no expert routing)
Vocabulary Size: 202,048

Input:

Input Type(s): Text, Image, Other (Video), Audio
Input Format(s): String, Red, Green, Blue (RGB), Other (Video), Other (Waveform)
Input Parameters: One-Dimensional (1D), Two-Dimensional (2D), Three-Dimensional (3D), One-Dimensional (1D)
Other Properties Related to Input: Requests support temperature, top_p, and top_k parameters. Combined input and output context length is up to 1,000,000 tokens.

Output:

Output Type(s): Text
Output Format: String
Output Parameters: One-Dimensional (1D)
Other Properties Related to Output: The combined input and output context length is up to 1,000,000 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:

Supported Runtime Engine(s):

TRT-LLM
vLLM
SGLang

Supported Hardware Microarchitecture Compatibility:

NVIDIA Blackwell

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):

v1.0: Muse-Spark-Open-816B-A42B-DFlash2 is the DFlash 2 draft model for Muse-Spark-Open-816B-A42B, trained with a pre-release build of NVIDIA Model Optimizer 0.47 and released in BF16 (bfloat16)

Training and Evaluation Datasets:

Training Dataset:

Link: Nemotron-Post-Training-Dataset-v2, only prompts from the datasets were used for data synthesis (the original responses from GPT were not used), which is then used to train the DFlash modules.
Data Modality: Text
Text Training Data Size: 1 Billion to 10 Trillion Tokens
Data Collection Method by dataset: Hybrid: Automated, Synthetic
Labeling Method by dataset: Hybrid: Automated, Synthetic
Properties: About 3.48 million multilingual text samples featuring prompts spanning math, code, STEM, and conversational topics. Each sample includes a synthetic response generated by the target model; the corpus mixes responses generated at five reasoning-effort levels, each including its reasoning trace.

Testing Dataset:

Link: SPEED-Bench, MT-Bench
Data Collection Method by dataset: Hybrid: Automated, Manually-Collected, Synthetic
Labeling Method by dataset: Hybrid: Automated, Manually-Labeled, Synthetic
Properties: The SPEED-Bench qualitative split (880 text prompts in 11 task categories, mostly English), the 32k-token bucket of the SPEED-Bench throughput split (first 100 prompts of each of its three categories), and MT-Bench (80 English two-turn text questions in 8 categories) were used during development to measure acceptance length and select checkpoints.

Evaluation Dataset:

The model is evaluated on the SPEED-Bench and MT-Bench datasets for speculative decoding benchmarking.

Data Collection Method by dataset: Hybrid: Automated, Manually-Collected, Synthetic
Labeling Method by dataset: Hybrid: Automated, Manually-Labeled, Synthetic
Properties: All prompts are text and mostly in English. The SPEED-Bench qualitative split has 880 prompts in 11 task categories (such as coding, math, reasoning, and writing, plus a multilingual category covering 23 languages); MT-Bench has 80 two-turn questions in 8 categories. Acceptance length over the target model's responses measures the quality of the draft model.

Inference:

Acceleration Engine: vLLM, SGLang
Test Hardware: NVIDIA Blackwell GB300

DFlash 2 Speculative Decoding

This model is ready for inference in DFlash 2 speculative decoding mode. DFlash 2 drafts a block of candidate tokens in a single non-causal forward pass, then uses a pairwise path selector over each position’s top-k candidates. The target model verifies the proposed path so that more than one token can be accepted per generation step. The mean number of tokens emitted per step is the acceptance length.

Usage

vLLM

vllm serve meta-models/Muse-Spark-Open-816B-A42B \
    --speculative-config '{"method": "dflash", "model": "nvidia/Muse-Spark-Open-816B-A42B-DFlash2", "num_speculative_tokens": 7, "draft_sample_method": "probabilistic"}' \
    --trust-remote-code

SGLang

Run inside the Muse Spark Open image, lmsysorg/sglang:dev-muse-spark-open. On 4 Blackwell GPUs, serve the NVFP4 target with this draft model:

python -m sglang.launch_server \
    --model-path meta-models/Muse-Spark-Open-816B-A42B-NVFP4 \
    --quantization modelopt_fp4 \
    --moe-runner-backend flashinfer_trtllm \
    --tp-size 4 \
    --ep-size 4 \
    --attention-backend trtllm_mha \
    --page-size 128 \
    --kv-cache-dtype fp8_e4m3 \
    --mem-fraction-static 0.8 \
    --reasoning-parser muse \
    --tool-call-parser muse \
    --enable-layernorm-sp \
    --enable-megamoe-for-layernorm-sp \
    --speculative-algorithm DFLASH \
    --speculative-draft-model-path nvidia/Muse-Spark-Open-816B-A42B-DFlash2 \
    --speculative-draft-attention-backend fa4 \
    --speculative-num-draft-tokens 8 \
    --host 0.0.0.0 \
    --port 30000

Evaluation

All acceptance lengths below were measured with reasoning effort xhigh, 4,096 output tokens, and concurrency 32 on one NVIDIA GB300 node (TP4 + EP4). The t=0 numbers use greedy decoding; the t=1 numbers use temperature 1.0 with no top-p or top-k truncation and draft_sample_method: probabilistic.

Acceptance length on MT-Bench with draft length 7, measured with vLLM and Muse-Spark-Open-816B-A42B (NVFP4):

Category	AL, greedy (t=0)	AL, sampling (t=1)
writing	3.94	3.42
roleplay	3.32	2.93
reasoning	3.53	3.29
math	4.81	4.24
coding	4.46	4.01
extraction	4.58	4.16
stem	3.35	3.05
humanities	2.96	2.79
Overall Average	3.87	3.49

Acceptance length on SPEED-Bench (qualitative split) with draft length 7, measured with vLLM and Muse-Spark-Open-816B-A42B (NVFP4):

Category	AL, greedy (t=0)	AL, sampling (t=1)
coding	4.29	3.89
humanities	3.27	2.92
math	4.08	3.63
multilingual	4.47	4.12
qa	3.52	3.24
rag	4.33	3.85
reasoning	3.99	3.53
roleplay	3.50	3.11
stem	3.47	3.16
summarization	4.17	3.84
writing	3.10	2.85
Overall Average	3.84	3.47

Acceptance length on SPEED-Bench (throughput split, 32k-token inputs, first 100 prompts of each category) with draft length 7, measured with vLLM and Muse-Spark-Open-816B-A42B (NVFP4):

Category	AL, greedy (t=0)	AL, sampling (t=1)
mixed	3.78	3.37
low_entropy	3.75	3.38
high_entropy	2.52	2.35
Overall Average	3.35	3.03

Model Limitations:

The base 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. The model may generate answers that may be inaccurate, omit key information, or include irrelevant or redundant text producing socially unacceptable or undesirable text, even if the prompt itself does not include anything explicitly offensive.

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, please see the Model Card++ Explainability, Bias, Safety & Security, and Privacy Subcards below.

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

SUBCARDS:

Explainability

Field:	Response:
Intended Task/Domain:	Text generation, reasoning, summarization, and question answering.
Model Type:	Text and Image-to-text transformer
Intended Users:	This model is intended for developers, researchers, and customers building/utilizing LLMs, while balancing accuracy and efficiency.
Output:	Text String(s)
Describe how the model works:	Generates text by predicting the next word or token based on the context provided in the input sequence using multiple self-attention layers
Name the adversely impacted groups this has been tested to deliver comparable outcomes regardless of:	Not Applicable
Technical Limitations & Mitigation:	The 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.
Verified to have met prescribed quality standards?	Yes
Performance Metrics:	Accuracy, Throughput, and user-side throughput
Potential Known Risk:	The model may generate answers that may be inaccurate, omit key information, or include irrelevant or redundant text producing socially unacceptable or undesirable text, even if the prompt itself does not include anything explicitly offensive.
Licensing:	GOVERNING TERMS: Use of the model/checkpoints is governed by the NVIDIA Open Model Agreement. Additional Information: Meta’s Muse Model License. Built with Muse Spark Open.

Bias

Field	Response
Participation considerations from adversely impacted groups 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, Chatbot Development, Code Generation, Reasoning
Describe life critical application (if present):	Not Applicable
Use Case Restrictions:	Abide by the GOVERNING TERMS: Use of the model/checkpoints is governed by the NVIDIA Open Model Agreement. Additional Information: Meta’s Muse Model License. Built with Muse Spark Open.
Model and Dataset Restrictions:	The Principle of least privilege (PoLP) is applied limiting access for dataset generation. Restrictions enforce dataset access during training, and dataset license constraints adhered to. Model checkpoints are made available on Hugging Face, and may become available on cloud providers' model catalog.

Privacy

Field	Response
Generatable or reverse-engineerable personal data	No
Was consent obtained for any personal data used	Not Applicable
Personal data used to create this model	None Known
How often is the dataset reviewed	Before Release
Was data from user interactions with the AI model used to train the model	No
Is there provenance for all datasets used in training	Yes
Does data labeling comply with privacy laws	Yes
Is the data compliant with data-subject requests for correction or removal	No, not possible with externally-sourced data.
Applicable NVIDIA Privacy Policy	NVIDIA Privacy Policy