---
license: mit
language:
- en
- ko
tags:
- KT
- K-intelligence
- Mi:dm
pipeline_tag: text-generation
library_name: transformers
---
Mi:dm 2.0-Base
π€ Mi:dm 2.0 Models |
π Mi:dm 2.0 Technical Report* |
π Mi:dm 2.0 Technical Blog*
*To be released soon
# News π’
- π _(Coming Soon!) GGUF format model files will be available soon for easier local deployment._
- β‘οΈ`2025/07/04`: Released Mi:dm 2.0 Model collection on Hugging Faceπ€.
# Table of Contents
- ___Overview___
- [Mi:dm 2.0](#midm-20)
- [Quickstart](#quickstart)
- [Evaluation](#evaluation)
- ___Usage___
- [Run on Friendli.AI](#run-on-friendliai)
- [Run on Your Local Machine](#run-on-your-local-machine)
- [Deployment](#deployment)
- [Tutorials](#tutorials)
- ___More Information___
- [Limitation](#limitation)
- [License](#license)
- [Contact](#contact)
# Overview
### Mi:dm 2.0
Mi:dm 2.0 is a __"Korean-centric AI"__ model developed with KT's proprietary technology. __"Korean-centric AI"__ refers to a model that thoroughly internalizes the unique values, cognitive frameworks, and commonsense reasoning intrinsic to Korean society. It is not simply about processing and responding in Korean; it is about the profound understanding that reflects and respects the socio-cultural fabric of Korean norms and values.
The newly introduced Mi:dm 2.0 model comes in two versions:
* **Mi:dm 2.0-Mini** is a 2.3B parameter Dense small model, designed for seamless use in environments such as on-device settings and low-end GPUs. It was created by pruning and distilling the Base model.
* **Mi:dm 2.0-Base** has 11.5B parameters and was designed to balance model size and performance by expanding an 8B scale model using the DuS (Depth-up Scaling) method. It's a practical model that can be applied to various real-world services, considering both performance and versatility.
> [!Note]
> Neither the pre-training nor the post-training data includes KT users' data.
### Quickstart
Here is the code snippet to run conversational inference with the model:
```python
import torch
from transformers import AutoModelForCausalLM, AutoTokenizer, GenerationConfig
model_name = "K-intelligence/Midm-2.0-Base-Instruct"
model = AutoModelForCausalLM.from_pretrained(
model_name,
torch_dtype=torch.bfloat16,
trust_remote_code=True,
device_map="auto"
)
tokenizer = AutoTokenizer.from_pretrained(model_name)
generation_config = GenerationConfig.from_pretrained(model_name)
prompt = "KTμ λν΄ μκ°ν΄μ€"
# message for inference
messages = [
{"role": "system",
"content": "Mi:dm(λ―Ώ:μ)μ KTμμ κ°λ°ν AI κΈ°λ° μ΄μμ€ν΄νΈμ΄λ€."},
{"role": "user", "content": prompt}
]
input_ids = tokenizer.apply_chat_template(
messages,
tokenize=True,
add_generation_prompt=True,
return_tensors="pt"
)
output = model.generate(
input_ids.to("cuda"),
generation_config=generation_config,
eos_token_id=tokenizer.eos_token_id,
max_new_tokens=128,
do_sample=False,
)
print(tokenizer.decode(output[0]))
```
> [!NOTE]
> The `transformers` library should be version `4.45.0` or higher.
# Evaluation
#### English
| Benchmark |
Exaone-3.5-2.4B-inst |
Qwen3-4B |
Mi:dm 2.0-Mini-inst |
Exaone-3.5-7.8B-inst |
Qwen3-14B |
Llama-3.1-8B-inst |
Mi:dm 2.0-Base-inst |
| Instruction Following |
IFEval |
81.1 |
79.7 |
73.6 |
83.6 |
83.9 |
79.9 |
84.0 |
| Reasoning |
BBH |
46.4 |
79.0 |
44.5 |
50.1 |
83.4 |
60.3 |
77.7 |
| GPQA |
28.1 |
39.8 |
26.6 |
33.1 |
49.8 |
21.6 |
33.5 |
| MuSR |
49.7 |
58.5 |
51.7 |
51.2 |
57.7 |
50.3 |
51.9 |
| Avg. |
41.4 |
59.1 |
40.9 |
44.8 |
63.6 |
44.1 |
54.4 |
| Mathematics |
GSM8K |
82.5 |
90.4 |
83.1 |
81.1 |
88.0 |
81.2 |
91.6 |
| MBPP+ |
59.8 |
62.4 |
60.9 |
79.4 |
73.4 |
81.8 |
77.5 |
| General Knowledge |
MMLU-pro |
- |
- |
- |
40.7 |
70.5 |
47.6 |
53.3 |
| MMLU |
59.5 |
73.3 |
56.5 |
69.0 |
82.7 |
70.7 |
73.7 |
| Avg. |
59.5 |
73.3 |
56.5 |
54.8 |
76.6 |
59.2 |
63.5 |
#### Korean
| Benchmark |
Exaone-3.5-2.4B-inst |
Qwen3-4B |
Mi:dm 2.0-Mini-inst |
Exaone-3.5-7.8B-inst |
Qwen3-14B |
Llama-3.1-8B-inst |
Mi:dm 2.0-Base-inst |
| Comprehension |
K-Prag* |
68.7 |
73.9 |
69.5 |
73.5 |
86.7 |
59.9 |
86.5 |
| K-Refer-Hard* |
58.5 |
56.7 |
55.4 |
61.9 |
74.0 |
48.6 |
70.8 |
| Ko-Best |
87.2 |
91.5 |
80.5 |
92.0 |
93.9 |
77.4 |
95.2 |
| Ko-Sovereign* |
38.0 |
43.5 |
42.5 |
44.0 |
52.0 |
31.5 |
53.0 |
| Avg. |
62.5 |
66.6 |
61.9 |
67.2 |
76.8 |
51.5 |
76.1 |
| Reasoning |
Ko-Winogrande |
60.3 |
67.5 |
61.7 |
64.6 |
77.2 |
40.1 |
75.1 |
| Ko-Best |
64.1 |
69.2 |
64.5 |
60.3 |
75.4 |
26.0 |
73.0 |
| LogicKor* |
7.4 |
5.6 |
7.7 |
8.6 |
6.4 |
2.4 |
8.6 |
| HRM8K* |
38.5 |
56.7 |
39.9 |
49.7 |
64.5 |
30.9 |
52.9 |
| Avg. |
36.7 |
43.8 |
37.4 |
39.5 |
48.8 |
19.8 |
44.8 |
| Society & Culture |
K-Refer* |
64.0 |
53.6 |
66.4 |
71.6 |
72.4 |
43.2 |
89.6 |
| K-Refer-Hard* |
67.1 |
42.9 |
61.4 |
69.3 |
65.7 |
36.4 |
86.4 |
| Ko-Sovereign* |
44.4 |
35.8 |
36.7 |
46.9 |
49.8 |
33.8 |
56.3 |
| HAERAE* |
61.3 |
50.6 |
70.8 |
72.9 |
68.4 |
49.5 |
81.5 |
| Avg. |
59.2 |
45.7 |
58.8 |
65.2 |
64.1 |
40.7 |
78.4 |
| Reasoning (Domain) |
KMMLU |
43.5 |
50.6 |
45.1 |
52.6 |
55.4 |
33.0 |
57.3 |
| Ko-Sovereign* |
42.4 |
42.5 |
42.4 |
45.6 |
54.7 |
36.7 |
58.0 |
| Avg. |
43.0 |
46.5 |
43.8 |
49.1 |
55.1 |
34.8 |
57.7 |
| Instruction Following |
Ko-IFEval* |
65.4 |
75.9 |
73.3 |
69.1 |
83.6 |
60.1 |
82.0 |
| Ko-MTBench |
74.0 |
63.0 |
74.0 |
79.6 |
71.0 |
57.0 |
89.7 |
| Avg. |
68.9 |
69.4 |
73.6 |
74.4 |
77.3 |
58.5 |
85.9 |
`*` indicates KT proprietary evaluation resources.
# Usage
### Run on Friendli.AI
You can try our model immediately via `Friendli.AI`. Simply click `Deploy` and then `Friendli Endpoints`.
> [!Note]
> Please note that a login to `Friendli.AI` is required after your fifth chat interaction.
### Run on Your Local Machine
We provide a detailed description about running Mi:dm 2.0 on your local machine using llama.cpp, LM Studio, and Ollama. Please check our [github]() for more information
### Deployment
To serve Mi:dm 2.0 using [vLLM](https://github.com/vllm-project/vllm)(`>=0.8.0`) with an OpenAI-compatible API:
```bash
vllm serve K-intelligence/Midm-2.0-Base-Instruct
```
### Tutorials
To help our end-users easily use Mi:dm 2.0, we have provided comprehensive tutorials on [github]().
# More Information
### Limitation
* The training data for both Mi:dm 2.0 models consists primarily of English and Korean. Understanding and generation in other languages are not guaranteed.
* The model is not guaranteed to provide reliable advice in fields that require professional expertise, such as law, medicine, or finance.
* Researchers have made efforts to exclude unethical content from the training data β such as profanity, slurs, bias, and discriminatory language. However, despite these efforts, the model may still produce inappropriate expressions or factual inaccuracies.
### License
Mi:dm 2.0 is licensed under the [MIT License](./LICENSE).
### Contact
- Mi:dm 2.0 Technical Inquiries: midm-llm@kt.com