I've been experimenting with a new quantization approach that selectively elevates the precision of key layers beyond what the default IMatrix configuration provides.
In my testing, standard IMatrix quantization underperforms at lower bit depths, especially with Mixture of Experts (MoE) models. To address this, I'm using the --tensor-type option in llama.cpp to manually "bump" important layers to higher precision. You can see the implementation here:
👉 Layer bumping with llama.cpp
While this does increase model file size, it significantly improves precision for a given quantization level.
I'd love your feedback—have you tried this? How does it perform for you?
Mi:dm 2.0 is a "Korea-centric AI" model developed using KT's proprietary technology. The term "Korea-centric AI" refers to a model that deeply internalizes the unique values, cognitive frameworks, and commonsense reasoning inherent to Korean society. It goes beyond simply processing or generating Korean text—it reflects a deeper understanding of the socio-cultural norms and values that define Korean society.
Mi:dm 2.0 is released in two versions:
Mi:dm 2.0 Base
An 11.5B parameter dense model designed to balance model size and performance.
It extends an 8B-scale model by applying the Depth-up Scaling (DuS) method, making it suitable for real-world applications that require both performance and versatility.
Mi:dm 2.0 Mini
A lightweight 2.3B parameter dense model optimized for on-device environments and systems with limited GPU resources.
It was derived from the Base model through pruning and distillation to enable compact deployment.
[!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
1import torch
2from transformers import AutoModelForCausalLM, AutoTokenizer, GenerationConfig
34model_name ="K-intelligence/Midm-2.0-Base-Instruct"56model = AutoModelForCausalLM.from_pretrained(7 model_name,8 torch_dtype=torch.bfloat16,9 trust_remote_code=True,10 device_map="auto"11)12tokenizer = AutoTokenizer.from_pretrained(model_name)13generation_config = GenerationConfig.from_pretrained(model_name)1415prompt ="KT에 대해 소개해줘"1617# message for inference18messages =[19{"role":"system",20"content":"Mi:dm(믿:음)은 KT에서 개발한 AI 기반 어시스턴트이다."},21{"role":"user","content": prompt}22]2324input_ids = tokenizer.apply_chat_template(25 messages,26 tokenize=True,27 add_generation_prompt=True,28 return_tensors="pt"29)3031output = model.generate(32 input_ids.to("cuda"),33 generation_config=generation_config,34 eos_token_id=tokenizer.eos_token_id,35 max_new_tokens=128,36 do_sample=False,37)38print(tokenizer.decode(output[0]))
[!NOTE]
The transformers library should be version 4.45.0 or higher.
Evaluation
Korean
Model
Society & Culture
General Knowledge
Instruction Following
K-Refer*
K-Refer-Hard*
Ko-Sovereign*
HAERAE
Avg.
KMMLU
Ko-Sovereign*
Avg.
Ko-IFEval
Ko-MTBench
Avg.
Qwen3-4B
53.6
42.9
35.8
50.6
45.7
50.6
42.5
46.5
75.9
63.0
69.4
Exaone-3.5-2.4B-inst
64.0
67.1
44.4
61.3
59.2
43.5
42.4
43.0
65.4
74.0
68.9
Mi:dm 2.0-Mini-inst
66.4
61.4
36.7
70.8
58.8
45.1
42.4
43.8
73.3
74.0
73.6
Qwen3-14B
72.4
65.7
49.8
68.4
64.1
55.4
54.7
55.1
83.6
71
77.3
Llama-3.1-8B-inst
43.2
36.4
33.8
49.5
40.7
33.0
36.7
34.8
60.1
57
58.5
Exaone-3.5-7.8B-inst
71.6
69.3
46.9
72.9
65.2
52.6
45.6
49.1
69.1
79.6
74.4
Mi:dm 2.0-Base-inst
89.6
86.4
56.3
81.5
78.4
57.3
58.0
57.7
82
89.7
85.9
Model
Comprehension
Reasoning
K-Prag*
K-Refer-Hard*
Ko-Best
Ko-Sovereign*
Avg.
Ko-Winogrande
Ko-Best
LogicKor
HRM8K
Avg.
Qwen3-4B
73.9
56.7
91.5
43.5
66.6
67.5
69.2
5.6
56.7
43.8
Exaone-3.5-2.4B-inst
68.7
58.5
87.2
38.0
62.5
60.3
64.1
7.4
38.5
36.7
Mi:dm 2.0-Mini-inst
69.5
55.4
80.5
42.5
61.9
61.7
64.5
7.7
39.9
37.4
Qwen3-14B
86.7
74.0
93.9
52.0
76.8
77.2
75.4
6.4
64.5
48.8
Llama-3.1-8B-inst
59.9
48.6
77.4
31.5
51.5
40.1
26.0
2.4
30.9
19.8
Exaone-3.5-7.8B-inst
73.5
61.9
92.0
44.0
67.2
64.6
60.3
8.6
49.7
39.5
Mi:dm 2.0-Base-inst
86.5
70.8
95.2
53.0
76.1
75.1
73.0
8.6
52.9
44.8
* indicates KT proprietary evaluation resources.
English
Model
Instruction
Reasoning
Math
Coding
General Knowledge
IFEval
BBH
GPQA
MuSR
Avg.
GSM8K
MBPP+
MMLU-pro
MMLU
Avg.
Qwen3-4B
79.7
79.0
39.8
58.5
59.1
90.4
62.4
-
73.3
73.3
Exaone-3.5-2.4B-inst
81.1
46.4
28.1
49.7
41.4
82.5
59.8
-
59.5
59.5
Mi:dm 2.0-Mini-inst
73.6
44.5
26.6
51.7
40.9
83.1
60.9
-
56.5
56.5
Qwen3-14B
83.9
83.4
49.8
57.7
63.6
88.0
73.4
70.5
82.7
76.6
Llama-3.1-8B-inst
79.9
60.3
21.6
50.3
44.1
81.2
81.8
47.6
70.7
59.2
Exaone-3.5-7.8B-inst
83.6
50.1
33.1
51.2
44.8
81.1
79.4
40.7
69.0
54.8
Mi:dm 2.0-Base-inst
84.0
77.7
33.5
51.9
54.4
91.6
77.5
53.3
73.7
63.5
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.
Left Image
Right Image
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(>=0.8.0) with an OpenAI-compatible API:
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 — such as profanity, slurs, bias, and discriminatory language. However, despite these efforts, the model may still produce inappropriate expressions or factual inaccuracies.
The full Open Source Code for the Quantum Network Monitor Service available at my github repos ( repos with NetworkMonitor in the name) : Source Code Quantum Network Monitor. You will also find the code I use to quantize the models if you want to do it yourself GGUFModelBuilder
💬 How to test:
Choose an AI assistant type:
TurboLLM (GPT-4.1-mini)
HugLLM (Hugginface Open-source models)
TestLLM (Experimental CPU-only)
What I’m Testing
I’m pushing the limits of small open-source models for AI network monitoring, specifically:
Function calling against live network services
How small can a model go while still handling:
Automated Nmap security scans
Quantum-readiness checks
Network Monitoring tasks
🟡 TestLLM – Current experimental model (llama.cpp on 2 CPU threads on huggingface docker space):
✅ Zero-configuration setup
⏳ 30s load time (slow inference but no API costs) . No token limited as the cost is low.
🔧 Help wanted! If you’re into edge-device AI, let’s collaborate!
Other Assistants
🟢 TurboLLM – Uses gpt-4.1-mini :
**It performs very well but unfortunatly OpenAI charges per token. For this reason tokens usage is limited.
Create custom cmd processors to run .net code on Quantum Network Monitor Agents
Real-time network diagnostics and monitoring
Security Audits
Penetration testing (Nmap/Metasploit)
🔵 HugLLM – Latest Open-source models:
🌐 Runs on Hugging Face Inference API. Performs pretty well using the lastest models hosted on Novita.
💡 Example commands you could test:
"Give me info on my websites SSL certificate"
"Check if my server is using quantum safe encyption for communication"
"Run a comprehensive security audit on my server"
'"Create a cmd processor to .. (what ever you want)" Note you need to install a Quantum Network Monitor Agent to run the .net code on. This is a very flexible and powerful feature. Use with caution!
Final Word
I fund the servers used to create these model files, run the Quantum Network Monitor service, and pay for inference from Novita and OpenAI—all out of my own pocket. All the code behind the model creation and the Quantum Network Monitor project is open source. Feel free to use whatever you find helpful.
If you appreciate the work, please consider buying me a coffee ☕. Your support helps cover service costs and allows me to raise token limits for everyone.
I'm also open to job opportunities or sponsorship.
