Hugging Face's logo

cyankiwi
/
Ministral-3-3B-Instruct-2512-AWQ-4bit

Safetensors
vllm
mistral3
mistral-common
compressed-tensors
Model card Files
xet
Community
cpatonn commited on
Commit
b94f9ad
·
verified ·
1 Parent(s): 15b3c52

Create README.md

Browse files
Files changed (1) hide show
  1. README.md +130 -0
README.md ADDED
@@ -0,0 +1,130 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ library_name: vllm
3
+ language:
4
+ - en
5
+ - fr
6
+ - es
7
+ - de
8
+ - it
9
+ - pt
10
+ - nl
11
+ - zh
12
+ - ja
13
+ - ko
14
+ - ar
15
+ license: apache-2.0
16
+ inference: false
17
+ base_model:
18
+ - mistralai/Ministral-3-3B-Instruct-2512
19
+ extra_gated_description: >-
20
+ If you want to learn more about how we process your personal data, please read
21
+ our <a href="https://mistral.ai/terms/">Privacy Policy</a>.
22
+ tags:
23
+ - mistral-common
24
+ ---
25
+
26
+ # Ministral 3 3B Instruct 2512 BF16
27
+
28
+ The smallest model in the Ministral 3 family, **Ministral 3 3B** is a powerful, efficient tiny language model with vision capabilities.
29
+
30
+ This model is the instruct post-trained version, fine-tuned for instruction tasks, making it ideal for chat and instruction based use cases.
31
+
32
+ The Ministral 3 family is designed for edge deployment, capable of running on a wide range of hardware. Ministral 3 3B can even be deployed locally, capable of fitting in 16GB of VRAM in BF16, and less than 8GB of RAM/VRAM when quantized.
33
+
34
+ We provide a no-loss FP8 version [here](https://huggingface.co/mistralai/Ministral-3-3B-Instruct-2512), you can find other formats and quantizations in the [Ministral 3 - Additional Checkpoints](https://huggingface.co/collections/mistralai/ministral-3-additional-checkpoints) collection.
35
+
36
+ ## Key Features
37
+ Ministral 3 3B consists of two main architectural components:
38
+ - **3.4B Language Model**
39
+ - **0.4B Vision Encoder**
40
+
41
+ The Ministral 3 3B Instruct model offers the following capabilities:
42
+ - **Vision**: Enables the model to analyze images and provide insights based on visual content, in addition to text.
43
+ - **Multilingual**: Supports dozens of languages, including English, French, Spanish, German, Italian, Portuguese, Dutch, Chinese, Japanese, Korean, Arabic.
44
+ - **System Prompt**: Maintains strong adherence and support for system prompts.
45
+ - **Agentic**: Offers best-in-class agentic capabilities with native function calling and JSON outputting.
46
+ - **Edge-Optimized**: Delivers best-in-class performance at a small scale, deployable anywhere.
47
+ - **Apache 2.0 License**: Open-source license allowing usage and modification for both commercial and non-commercial purposes.
48
+ - **Large Context Window**: Supports a 256k context window.
49
+
50
+ ### Use Cases
51
+ Ideal for lightweight, real-time applications on edge or low-resource devices, such as:
52
+ - Image captioning
53
+ - Text classification
54
+ - Real-time efficient translation
55
+ - Data extraction
56
+ - Short content generation
57
+ - Fine-tuning and specialization
58
+ - And more...
59
+
60
+ Bringing advanced AI capabilities to edge and distributed environments for embedded systems.
61
+
62
+ ## Ministral 3 Family
63
+
64
+ | Model Name | Type | Precision | Link |
65
+ |--------------------------------|--------------------|-----------|------------------------------------------------------------------------------------------|
66
+ | Ministral 3 3B Base 2512 | Base pre-trained | BF16 | [Hugging Face](https://huggingface.co/mistralai/Ministral-3-3B-Base-2512) |
67
+ | **Ministral 3 3B Instruct 2512** | **Instruct post-trained** | **BF16** | [Hugging Face](https://huggingface.co/mistralai/Ministral-3-3B-Instruct-2512) |
68
+ | Ministral 3 3B Reasoning 2512 | Reasoning capable | BF16 | [Hugging Face](https://huggingface.co/mistralai/Ministral-3-3B-Reasoning-2512) |
69
+ | Ministral 3 8B Base 2512 | Base pre-trained | BF16 | [Hugging Face](https://huggingface.co/mistralai/Ministral-3-8B-Base-2512) |
70
+ | Ministral 3 8B Instruct 2512 | Instruct post-trained | BF16 | [Hugging Face](https://huggingface.co/mistralai/Ministral-3-8B-Instruct-2512) |
71
+ | Ministral 3 8B Reasoning 2512 | Reasoning capable | BF16 | [Hugging Face](https://huggingface.co/mistralai/Ministral-3-8B-Reasoning-2512) |
72
+ | Ministral 3 14B Base 2512 | Base pre-trained | BF16 | [Hugging Face](https://huggingface.co/mistralai/Ministral-3-14B-Base-2512) |
73
+ | Ministral 3 14B Instruct 2512 | Instruct post-trained | BF16 | [Hugging Face](https://huggingface.co/mistralai/Ministral-3-14B-Instruct-2512) |
74
+ | Ministral 3 14B Reasoning 2512 | Reasoning capable | BF16 | [Hugging Face](https://huggingface.co/mistralai/Ministral-3-14B-Reasoning-2512) |
75
+
76
+ Other formats available [here](https://huggingface.co/collections/mistralai/ministral-3-additional-checkpoints).
77
+
78
+ ## Benchmark Results
79
+
80
+ We compare Ministral 3 to similar sized models.
81
+
82
+ ### Reasoning
83
+
84
+ | Model | AIME25 | AIME24 | GPQA Diamond | LiveCodeBench |
85
+ |---------------------------|-------------|-------------|--------------|---------------|
86
+ | **Ministral 3 14B** | <u>0.850</u>| <u>0.898</u>| <u>0.712</u> | <u>0.646</u> |
87
+ | Qwen3-14B (Thinking) | 0.737 | 0.837 | 0.663 | 0.593 |
88
+ | | | | | |
89
+ | **Ministral 3 8B** | 0.787 | <u>0.860</u>| 0.668 | <u>0.616</u> |
90
+ | Qwen3-VL-8B-Thinking | <u>0.798</u>| <u>0.860</u>| <u>0.671</u> | 0.580 |
91
+ | | | | | |
92
+ | **Ministral 3 3B** | <u>0.721</u>| <u>0.775</u>| 0.534 | <u>0.548</u> |
93
+ | Qwen3-VL-4B-Thinking | 0.697 | 0.729 | <u>0.601</u> | 0.513 |
94
+
95
+ ### Instruct
96
+
97
+ | Model | Arena Hard | WildBench | MATH Maj@1 | MM MTBench |
98
+ |---------------------------|-------------|------------|-------------|------------------|
99
+ | **Ministral 3 14B** | <u>0.551</u>| <u>68.5</u>| <u>0.904</u>| <u>8.49</u> |
100
+ | Qwen3 14B (Non-Thinking) | 0.427 | 65.1 | 0.870 | NOT MULTIMODAL |
101
+ | Gemma3-12B-Instruct | 0.436 | 63.2 | 0.854 | 6.70 |
102
+ | | | | | |
103
+ | **Ministral 3 8B** | 0.509 | <u>66.8</u>| 0.876 | <u>8.08</u> |
104
+ | Qwen3-VL-8B-Instruct | <u>0.528</u>| 66.3 | <u>0.946</u>| 8.00 |
105
+ | | | | | |
106
+ | **Ministral 3 3B** | 0.305 | <u>56.8</u>| 0.830 | 7.83 |
107
+ | Qwen3-VL-4B-Instruct | <u>0.438</u>| <u>56.8</u>| <u>0.900</u>| <u>8.01</u> |
108
+ | Qwen3-VL-2B-Instruct | 0.163 | 42.2 | 0.786 | 6.36 |
109
+ | Gemma3-4B-Instruct | 0.318 | 49.1 | 0.759 | 5.23 |
110
+
111
+ ### Base
112
+
113
+ | Model | Multilingual MMLU | MATH CoT 2-Shot | AGIEval 5-shot | MMLU Redux 5-shot | MMLU 5-shot | TriviaQA 5-shot |
114
+ |---------------------|-------------------|-----------------|----------------|-------------------|-------------|-----------------|
115
+ | **Ministral 3 14B** | 0.742 | <u>0.676</u> | 0.648 | 0.820 | 0.794 | 0.749 |
116
+ | Qwen3 14B Base | <u>0.754</u> | 0.620 | <u>0.661</u> | <u>0.837</u> | <u>0.804</u>| 0.703 |
117
+ | Gemma 3 12B Base | 0.690 | 0.487 | 0.587 | 0.766 | 0.745 | <u>0.788</u> |
118
+ | | | | | | | |
119
+ | **Ministral 3 8B** | <u>0.706</u> | <u>0.626</u> | 0.591 | 0.793 | <u>0.761</u>| <u>0.681</u> |
120
+ | Qwen 3 8B Base | 0.700 | 0.576 | <u>0.596</u> | <u>0.794</u> | 0.760 | 0.639 |
121
+ | | | | | | | |
122
+ | **Ministral 3 3B** | 0.652 | <u>0.601</u> | 0.511 | 0.735 | 0.707 | 0.592 |
123
+ | Qwen 3 4B Base | <u>0.677</u> | 0.405 | <u>0.570</u> | <u>0.759</u> | <u>0.713</u>| 0.530 |
124
+ | Gemma 3 4B Base | 0.516 | 0.294 | 0.430 | 0.626 | 0.589 | <u>0.640</u> |
125
+
126
+ ## License
127
+
128
+ This model is licensed under the [Apache 2.0 License](https://www.apache.org/licenses/LICENSE-2.0.txt).
129
+
130
+ *You must not use this model in a manner that infringes, misappropriates, or otherwise violates any third party’s rights, including intellectual property rights.*