nvidia/MiniMax-M2.7-DFlash

Model Overview

Description:

The NVIDIA MiniMax-M2.7 DFlash model is the DFlash draft head of MiniMax AI's MiniMax-M2.7 model, which is an auto-regressive language model that uses an optimized transformer architecture. For more information, please check here. The NVIDIA MiniMax-M2.7 DFlash model incorporates DFlash speculative decoding with Model Optimizer.

This model is for demonstration purposes and not for production usage.

License/Terms of Use:

Governing Terms: Use of this model is governed by the NVIDIA Software and Model Evaluation license.

ADDITIONAL INFORMATION : Non-Commercial MiniMax License. Copyright (c) 2026 MiniMax

Deployment Geography:

Global

Use Case:

Developers designing AI Agent systems, chatbots, RAG systems, and other AI-powered applications. Also suitable for typical instruction-following tasks.

Release Date:

Hugging Face 06/30/2026 via [https://huggingface.co/nvidia/MiniMax-M2.7-DFlash]

Reference(s):

• MiniMax-M2.7 release notes
• NVIDIA TensorRT Model Optimizer — Speculative Decoding (DFlash implementation)
• DFlash: Block Diffusion for Flash Speculative Decoding

Model Architecture:

Architecture Type: Transformers

Network Architecture: MiniMax M2 (MoE)

This model was developed based on MiniMaxAI/MiniMax-M2.7. Number of model parameters 1.31B (DFlash draft module, including the token embedding and LM head over the target vocabulary)

Input:

Input Type(s): Text

Input Format(s): String

Input Parameters: 1D (One Dimensional): Sequences

Other Properties Related to Input: Context length 196608 (trained at 4096; extended at export via YaRN rope_scaling)

Output:

Output Type(s): Text

Output Format: String

Output Parameters: 1D (One Dimensional): Sequences

Other Properties Related to Output: Maximum context length 196608 (the draft exports a YaRN rope_scaling to match the MiniMax-M2.7 target's context window)

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):

• vLLM

Supported Hardware Microarchitecture Compatibility:

• NVIDIA Blackwell

• NVIDIA Hopper

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 (v1) is trained with nvidia-modelopt 0.44.0**

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 DFlash modules. Data Modality: Text

Text Training Data Size: [Less than a Billion Tokens] Data Collection Method by dataset: [Hybrid: Automated, Synthetic] Labeling Method by dataset: [Hybrid: Automated, Synthetic] Properties: 469,568 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 Data Collection Method by datasetHybrid: Human, Synthetic Labeling Method by dataset:Hybrid: Human, Synthetic Properties: 3,300 multi-turn dialogue sequences, each annotated with expert preference votes.

Inference:

Acceleration Engine: vLLM

Test Hardware: NVIDIA Hopper (H100)

DFlash Speculative Decoding

Synthesized data was obtained from MiniMax AI's MiniMax-M2.7 model, which is then used to finetune the DFlash modules. This model is ready for inference with vLLM in DFlash speculative decoding mode. DFlash modules are used to predict candidate tokens beyond the next token. In the generation step, each forward DFlash module generates a distribution of tokens beyond the previous. The longest accepted candidate sequence is selected so that more than 1 token is returned in the generation step. The mean number of tokens emitted per step is the acceptance length.

Usage

To serve the checkpoint with vLLM in DFlash speculative-decoding mode:

vllm serve \
--speculative-config '{"method": "dflash", "model": "", "num_speculative_tokens": 7}' \
--tensor-parallel-size 4 \
--max-model-len 8192 \
--trust-remote-code

num_speculative_tokens is the draft length (the DFlash block_size is 8, so up to 7 speculative tokens per step). Lower it to trade peak acceptance length for lower per-step draft cost.

Evaluation

We report acceptance length (AL) — the mean number of tokens emitted per decoding step (the accepted draft tokens plus the one bonus token from the target model). Higher is better; with draft length _d_ the maximum is _d_+1.

Measurement setup: measured with the specdec_bench harness on vLLM DFlash speculative decoding, greedy decoding (temperature 0), 4096-token generations, target MiniMaxAI/MiniMax-M2.7, on NVIDIA H100. We list two draft lengths — 3 (for comparison with other published speculative-decoding cards) and 7 (DFlash block_size − 1, the recommended serving configuration). The released draft ships with a YaRN rope_scaling in its config so it drafts correctly at the target's full context length (see [Long-context](#long-context-32k) below); the acceptance...