yeyu-nvidia commited on
Commit
506ca1e
·
verified ·
1 Parent(s): 0a1ee27

Update model card to approved version (MR !62)

Browse files

Reflect the approved MiniMax-M2.7-DFlash model-card review: link the DFlash paper
(arXiv:2602.06036) instead of 'forthcoming'; advertise the shipped YaRN-extended 196608
context on Input/Output; plus the reviewed metadata fields (Bias Metric, verified-status rows).

Signed-off-by: Ye Yu <yeyu@nvidia.com>

Files changed (1) hide show
  1. README.md +49 -49
README.md CHANGED
@@ -18,15 +18,13 @@ tags:
18
  ## Description:
19
  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](https://huggingface.co/MiniMaxAI/MiniMax-M2.7). The NVIDIA MiniMax-M2.7 DFlash model incorporates DFlash speculative decoding with Model Optimizer.
20
 
21
- This model is ready for commercial/non-commercial use. <br>
22
 
23
- ## Third-Party Community Consideration
24
- This model is not owned or developed by NVIDIA. This model has been developed and built to a third-party's requirements for this application and use case; see link to Non-NVIDIA [(MiniMax-M2.7) Model Card](https://huggingface.co/MiniMaxAI/MiniMax-M2.7).
25
 
26
  ### License/Terms of Use:
27
- Governing Terms: Use of this model is governed by the [NVIDIA Open Model License](https://www.nvidia.com/en-us/agreements/enterprise-software/nvidia-open-model-license/) .
28
 
29
- **ADDITIONAL INFORMATION** : [Modified MIT License](https://huggingface.co/MiniMaxAI/MiniMax-M2.7/blob/main/LICENSE). **MiniMax M2.7** .
30
 
31
 
32
  ## Deployment Geography:
@@ -36,12 +34,13 @@ Global
36
  Developers designing AI Agent systems, chatbots, RAG systems, and other AI-powered applications. Also suitable for typical instruction-following tasks.
37
 
38
  ## Release Date:
39
- Hugging Face 06/11/2026 via [https://huggingface.co/nvidia/MiniMax-M2.7-DFlash]
40
 
41
  # Reference(s):
42
 
43
  - [MiniMax-M2.7 release notes](https://huggingface.co/MiniMaxAI/MiniMax-M2.7)
44
- - [NVIDIA TensorRT Model Optimizer — Speculative Decoding](https://github.com/NVIDIA/Model-Optimizer/tree/main/examples/speculative_decoding) (DFlash implementation; reference paper forthcoming)
 
45
 
46
  ## Model Architecture:
47
  **Architecture Type:** Transformers <br>
@@ -53,13 +52,13 @@ Hugging Face 06/11/2026 via [https://huggingface.co/nvidia/MiniMax-M2.7-DFlash]
53
  **Input Type(s):** Text <br>
54
  **Input Format(s):** String <br>
55
  **Input Parameters:** 1D (One Dimensional): Sequences <br>
56
- **Other Properties Related to Input:** Context length 4096
57
 
58
  ## Output:
59
  **Output Type(s):** Text <br>
60
  **Output Format:** String <br>
61
  **Output Parameters:** 1D (One Dimensional): Sequences <br>
62
- **Other Properties Related to Output:** None
63
 
64
  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. <br>
65
 
@@ -68,8 +67,8 @@ Our AI models are designed and/or optimized to run on NVIDIA GPU-accelerated sys
68
  * vLLM <br>
69
 
70
  **Supported Hardware Microarchitecture Compatibility:** <br>
71
- * NVIDIA Hopper <br>
72
  * NVIDIA Blackwell <br>
 
73
 
74
  **Preferred Operating System(s):** <br>
75
  * Linux <br>
@@ -97,7 +96,7 @@ The integration of foundation and fine-tuned models into AI systems requires add
97
 
98
  ## Inference:
99
  **Acceleration Engine:** vLLM <br>
100
- **Test Hardware:** NVIDIA H100 <br>
101
 
102
  ## DFlash Speculative Decoding
103
  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.
@@ -168,7 +167,7 @@ NVIDIA believes Trustworthy AI is a shared responsibility and we have establishe
168
 
169
  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.
170
 
171
- For more detailed information on ethical considerations for this model, please see the [Model Card++ Bias, Explainability, Safety & Security, and Privacy Subcards](https://gitlab-master.nvidia.com/api-catalog/examples).
172
 
173
  Please report model quality, risk, security vulnerabilities or NVIDIA AI Concerns [here](https://www.nvidia.com/en-us/support/submit-security-vulnerability/). <br>
174
 
@@ -176,49 +175,50 @@ SUBCARDS:
176
 
177
  # **Explainability**
178
 
179
- | Field | Response |
180
- |:----------------------------------------:|:-------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------:|
181
- | **Intended Task/Domain:** | Text generation, reasoning, summarization, and question answering. |
182
- | **Model Type:** | Text-to-text transformer (Mixture of Experts) |
183
- | **Intended Users:** | This model is intended for developers, researchers, and customers building/utilizing LLMs, while balancing accuracy and efficiency. |
184
- | **Output:** | Text String(s) |
185
- | **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. |
186
- | **Name the adversely impacted groups this has been tested to deliver comparable outcomes regardless of:** | Not Applicable |
187
- | **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. |
188
- | **Verified to have met prescribed NVIDIA quality standards:** | Yes |
189
- | **Performance Metrics:** | Accuracy, Throughput, and user-side throughput |
190
- | **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. |
191
- | **Licensing:** | Use of this model is governed by the [NVIDIA Open Model License](https://www.nvidia.com/en-us/agreements/enterprise-software/nvidia-open-model-license/). ADDITIONAL INFORMATION: [MiniMax-M2.7 License](https://huggingface.co/MiniMaxAI/MiniMax-M2.7/blob/main/LICENSE). Built with MiniMax-M2.7. |
192
-
 
193
 
194
  # **Bias**
195
 
196
- | Field | Response |
197
- |:---------------------------------------------------------------------------------------------------------------:|:--------:|
198
- | Participation considerations from adversely impacted groups [protected classes](https://www.senate.ca.gov/protected-classes) in model design and testing: | None |
199
- | Measures taken to mitigate against unwanted bias | None |
200
-
201
 
202
  # **Safety & Security**
203
 
204
- | Field | Response |
205
- |:-----------------------------------:|:------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------:|
206
- | Model Application Field(s): | Chat, Instruction Following, Chatbot Development, Code Generation, Reasoning |
207
- | Describe the life critical impact (if present) | Not Applicable |
208
- | Use Case Restrictions: | Abide by the [NVIDIA Open Model License](https://www.nvidia.com/en-us/agreements/enterprise-software/nvidia-open-model-license/). ADDITIONAL INFORMATION: [MiniMax-M2.7 License](https://huggingface.co/MiniMaxAI/MiniMax-M2.7/blob/main/LICENSE). Built with MiniMax-M2.7. |
209
- | 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. |
210
-
211
 
212
  # **Privacy**
213
 
214
- | Field | Response |
215
- |:-----------------------------------------------------------------------------------------------------:|:-----------------------------------------------------------------------------------:|
216
- | Generatable or reverse engineerable personal data? | No |
217
- | Was consent obtained for any personal data used? | Not Applicable |
218
- | Personal data used to create this model? | None Known |
219
- | How often is dataset reviewed? | Before Release |
220
- | Was data from user interactions with the AI model (e.g. user input and prompts) used to train the model? | No |
221
- | Is there provenance for all datasets used in training? | Yes |
222
- | Does data labeling (annotation, metadata) comply with privacy laws? | Yes |
223
- | Is data compliant with data subject requests for data correction or removal, if such a request was made? | No, not possible with externally-sourced data. |
224
- | Applicable NVIDIA Privacy Policy | https://www.nvidia.com/en-us/about-nvidia/privacy-policy/ |
 
18
  ## Description:
19
  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](https://huggingface.co/MiniMaxAI/MiniMax-M2.7). The NVIDIA MiniMax-M2.7 DFlash model incorporates DFlash speculative decoding with Model Optimizer.
20
 
21
+ This model is for demonstration purposes and not for production usage. <br>
22
 
 
 
23
 
24
  ### License/Terms of Use:
25
+ **Governing Terms:** Use of this model is governed by the [NVIDIA Software and Model Evaluation license](https://www.nvidia.com/en-us/agreements/enterprise-software/nvidia-software-and-model-evaluation-license/).
26
 
27
+ **ADDITIONAL INFORMATION** : [Non-Commercial MiniMax License](https://github.com/MiniMax-AI/MiniMax-M2.7/blob/main/LICENSE). **Copyright (c) 2026 MiniMax**
28
 
29
 
30
  ## Deployment Geography:
 
34
  Developers designing AI Agent systems, chatbots, RAG systems, and other AI-powered applications. Also suitable for typical instruction-following tasks.
35
 
36
  ## Release Date:
37
+ Hugging Face 06/30/2026 via [https://huggingface.co/nvidia/MiniMax-M2.7-DFlash]
38
 
39
  # Reference(s):
40
 
41
  - [MiniMax-M2.7 release notes](https://huggingface.co/MiniMaxAI/MiniMax-M2.7)
42
+ - [NVIDIA TensorRT Model Optimizer — Speculative Decoding](https://github.com/NVIDIA/Model-Optimizer/tree/main/examples/speculative_decoding) (DFlash implementation)
43
+ - [DFlash: Block Diffusion for Flash Speculative Decoding](https://arxiv.org/abs/2602.06036)
44
 
45
  ## Model Architecture:
46
  **Architecture Type:** Transformers <br>
 
52
  **Input Type(s):** Text <br>
53
  **Input Format(s):** String <br>
54
  **Input Parameters:** 1D (One Dimensional): Sequences <br>
55
+ **Other Properties Related to Input:** Context length 196608 (trained at 4096; extended at export via YaRN rope_scaling)
56
 
57
  ## Output:
58
  **Output Type(s):** Text <br>
59
  **Output Format:** String <br>
60
  **Output Parameters:** 1D (One Dimensional): Sequences <br>
61
+ **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)
62
 
63
  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. <br>
64
 
 
67
  * vLLM <br>
68
 
69
  **Supported Hardware Microarchitecture Compatibility:** <br>
 
70
  * NVIDIA Blackwell <br>
71
+ * NVIDIA Hopper <br>
72
 
73
  **Preferred Operating System(s):** <br>
74
  * Linux <br>
 
96
 
97
  ## Inference:
98
  **Acceleration Engine:** vLLM <br>
99
+ **Test Hardware:** NVIDIA Hopper (H100) <br>
100
 
101
  ## DFlash Speculative Decoding
102
  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.
 
167
 
168
  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.
169
 
170
+ For more detailed information on ethical considerations for this model, please see the Model Card++ Bias, Explainability, Safety & Security, and Privacy Subcards below.
171
 
172
  Please report model quality, risk, security vulnerabilities or NVIDIA AI Concerns [here](https://www.nvidia.com/en-us/support/submit-security-vulnerability/). <br>
173
 
 
175
 
176
  # **Explainability**
177
 
178
+ |Field:|Response:|
179
+ |:---:|:---:|
180
+ |Intended Task/Domain:| Text generation, reasoning, summarization, and question answering. |
181
+ |Model Type: |Text and Image-to-text transformer |
182
+ |Intended Users:|This model is intended for developers, researchers, and customers building/utilizing LLMs, while balancing accuracy and efficiency.|
183
+ |Output:|Text String(s)|
184
+ |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|
185
+ |Name the adversely impacted groups this has been tested to deliver comparable outcomes regardless of: | Not Applicable|
186
+ |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.|
187
+ |Verified to have met prescribed quality standards?|Yes|
188
+ |Performance Metrics:|Accuracy, Throughput, and user-side throughput|
189
+ |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. |
190
+ |Licensing:| Your usage is governed by the following **Governing Terms:** Use of this model is governed by the [NVIDIA Software and Model Evaluation license](https://www.nvidia.com/en-us/agreements/enterprise-software/nvidia-software-and-model-evaluation-license/).
191
+
192
+ **ADDITIONAL INFORMATION** : [Non-Commercial MiniMax License](https://github.com/MiniMax-AI/MiniMax-M2.7/blob/main/LICENSE). **Copyright (c) 2026 MiniMax** |
193
 
194
  # **Bias**
195
 
196
+ |Field:|Response:|
197
+ |:---:|:---:|
198
+ |Participation considerations from adversely impacted groups [protected classes](https://www.senate.ca.gov/content/protected-classes) in model design and testing:|Not Applicable|
199
+ |Measures taken to mitigate against unwanted bias:|N/A|
200
+ |Bias Metric:|Not Applicable|
201
 
202
  # **Safety & Security**
203
 
204
+ |Field:|Response:|
205
+ |:---:|:---:|
206
+ |Model Application(s):|Chat, Instruction Following, Chatbot Development, Code Generation, Reasoning|
207
+ |Describe life critical application (if present):|Not Applicable |
208
+ |Use Case Restrictions:|Abide by the **Governing Terms:** Use of this model is governed by the [NVIDIA Software and Model Evaluation license](https://www.nvidia.com/en-us/agreements/enterprise-software/nvidia-software-and-model-evaluation-license/).
209
+ **ADDITIONAL INFORMATION** : [Non-Commercial MiniMax License](https://github.com/MiniMax-AI/MiniMax-M2.7/blob/main/LICENSE). **Copyright (c) 2026 MiniMax** |
210
+ |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.|
211
 
212
  # **Privacy**
213
 
214
+ |Field:|Response:|
215
+ |:---:|:---:|
216
+ |Generatable or Reverse engineerable personal data?|No|
217
+ |Personal data used to create this model?|No|
218
+ |Was consent obtained for any personal data used?|Not Applicable|
219
+ |How often is dataset reviewed?|Before Release|
220
+ |Was data from user interactions with the AI model (e.g. user input and prompts) used to train the model? | No |
221
+ |Is there provenance for all datasets used in training?|Yes|
222
+ |Does data labeling (annotation, metadata) comply with privacy laws?|Not Applicable|
223
+ |Is data compliant with data subject requests for data correction or removal, if such a request was made? | Not Applicable|
224
+ |Applicable NVIDIA Privacy Policy|https://www.nvidia.com/en-us/about-nvidia/privacy-policy/|