Model Overview

Description:

The NVIDIA MiniMax-M3 DSpark model is the DSpark draft head of MiniMax's MiniMax-M3 model, which is a native multimodal mixture-of-experts language model that uses an optimized transformer architecture. For more information, please check here. The NVIDIA MiniMax-M3 DSpark model incorporates DSpark speculative decoding with Model Optimizer.

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

License/Terms of Use:

Governing Terms: Use of this model is governed by the NVIDIA Open Model License.

ADDITIONAL INFORMATION : MiniMax Community License. MiniMax-M3 .

Deployment Geography:

Global

Use Case:

Developers and inference providers building multimodal AI agents, coding assistants, long-context reasoning systems, and cowork-style assistants that benefit from lower-latency speculative decoding. This model is intended for DSpark-assisted serving of MiniMax-M3 for text, image, and video understanding workflows.

Release Date:

Hugging Face 07/22/2026 via https://huggingface.co/nvidia/MiniMax-M3-DSpark

References

• Nvidia Model Optimizer: https://github.com/NVIDIA/Model-Optimizer
• MiniMax-M3 base model card
• MiniMax Sparse Attention technical report
• DSpark: Confidence-Scheduled Speculative Decoding with Semi-Autoregressive Generation

Model Architecture:

Architecture Type: Transformers Network Architecture: MiniMaxM3SparseForConditionalGeneration Number of Model Parameters: 428B in total and 23B activated

Input:

Input Type(s): Text Input Format(s): String Input Parameters: One-Dimensional (1D), Two-Dimensional (2D), Three-Dimensional (3D) Other Properties Related to Input: Context length up to 1,048,576 tokens

Output:

Output Type(s): Text Output Format: String Output Parameters: One Dimensional(1D): Sequences Other Properties Related to Output: Outputs may include natural-language responses, code, structured reasoning, tool-use content, and multimodal analysis depending on application-level tooling.

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

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 DSpark version and is trained with nvidia-modelopt v0.45.0

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 DSpark modules. Data Modality: Text, Image, Video Image Training Data Size: Undisclosed Text Training Data Size: [1 Billion to 10 Trillion Tokens] Video Training Data Size: Undisclosed Data Collection Method by dataset: Hybrid: Automated, Synthetic Labeling Method by dataset: Hybrid: Automated, Synthetic Properties: 2M multilingual text samples featuring prompts spanning math, code, STEM, and conversational topics. Each sample includes a synthetic response generated by the target model.

Evaluation Dataset:

Link: MTBench, for more details, see here; SPEED-Bench Data Collection Method by dataset: Hybrid: Human, Synthetic Labeling Method by dataset:Hybrid: Human, Synthetic Properties: MT-Bench contains 3,300 multi-turn dialogue sequences, each annotated with expert preference votes. SPEED-Bench is a unified, diverse benchmark for speculative decoding spanning coding, humanities, math, multilingual, QA, RAG, reasoning, roleplay, STEM, summarization, and writing domains.

Inference:

Acceleration Engine: vLLM Test Hardware: NVIDIA B300

DSpark Speculative Decoding

Synthesized data was obtained from MiniMax's MiniMax-M3 model, which is then used to finetune the DSpark modules. This model is ready for inference with vLLM in DSpark speculative decoding mode, where a parallel draft backbone proposes a block of candidate tokens, a lightweight Markov head injects intra-block token dependency, and a confidence head predicts per-position acceptance probability for scheduled verification. Compared with a purely parallel drafter, DSpark is designed to improve accepted length while preserving the latency benefits of block drafting. The longest accepted candidate sequence is selected so that more than 1 token is returned in the generation step. The number of tokens generated in each step is called acceptance rate.

Usage

To serve the checkpoint with vLLM:

vllm serve MiniMaxAI/MiniMax-M3 \
  --tensor-parallel-size 4 \
  --block-size 128 \
  --trust-remote-code \
  --speculative-config '{
    "method": "dspark",
    "model": "<draft-model>",
    "num_speculative_tokens":8
  }'

Alternatively, with the Python LLM API:

from vllm import LLM, SamplingParams

llm = LLM(
    model="MiniMaxAI/MiniMax-M3",
    tensor_parallel_size=4,
    block_size=128,
    trust_remote_code=True,
    speculative_config={
        "method": "dspark",
        "model": "<draft-model>",
        "num_speculative_tokens":8
    },
)

Note: --block-size 128 (the KV-cache block size) is required by MiniMax-M3's sparse attention.

Training

This checkpoint was trained in two phases: short-context training (4k sequence length) on M3-synthesized Speculative-Decoding-Dataset-v2, followed by a long-context continuation (16k sequence length) on a 12k+-token subset of the same corpus, which improves long-context acceptance (throughput-32k) while preserving short-context quality.

See the following streaming recipe for NVIDIA Model Optimizer as a reference for training your own DSpark draft head:

uv run launch.py --yaml examples/MiniMaxAI/MiniMax-M3/hf_streaming_dspark_multi_node.yaml --yes

Run from tools/launcher; see its README for setup.

Evaluation

Acceptance rate on SPEED-Bench (qualitative subset) with a draft block size of 8:

Acceptance rate on SPEED-Bench (throughput-32k subset, long-context) with a draft block size of 8:

Acceptance rate on MT-bench with a draft block size of 8:

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.

Please make sure you have proper rights and permissions for all input image and video content; if image or video includes people, personal health information, or intellectual property, the image or video generated will not blur or maintain proportions of image subjects included.

For more detailed information on ethical considerations for this model, please 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

ADDITIONAL INFORMATION : MiniMax Community License. MiniMax-M3 . |

Bias

Safety & Security

ADDITIONAL INFORMATION : MiniMax Community License. MiniMax-M3 . |
|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