Mi:dm 2.0
Korea-centric language model by KT
KT released Mi:dm 2.0 in July 2025 as a language model focused on Korean cultural values and reasoning. It comes in two sizes, an 11.5 billion parameter version and a smaller 2.3 billion variant for local use.
- Released by KT in July 2025
- Available in 11.5B and 2.3B sizes
- Licensed under MIT terms
- Focuses on Korean cultural values
- Supports local deployment via Ollama
What it is
Mi:dm 2.0 is a language model built by the South Korean telecommunications company KT. It launched in July 2025 with a focus on understanding Korean social norms and cognitive frameworks. The system is available in two dense configurations. The larger version holds 11.5 billion parameters, while the smaller one uses 2.3 billion parameters for devices with limited graphics processing power. Both models process text in English and Korean.
How it came about
KT published a technical blog post about the architecture in August 2025. The full model collection appeared on Hugging Face later that summer. Support for function calling via vLLM arrived in October 2025. The developers state that no user data from KT services was used during pre-training or post-training phases.
Inside the model
The 11.5 billion parameter model uses a method called Depth-up Scaling to extend an 8 billion parameter base. This approach aims to balance size with performance for general tasks. The smaller 2.3 billion variant underwent pruning and distillation to reduce its footprint for local deployment. Training data consisted mainly of English and Korean texts. The team worked to remove unethical content such as slurs and bias from the dataset.
What it can do
The model is designed to reflect Korean commonsense reasoning and socio-cultural values. It supports function calling when used with a specific parser tool. Performance in languages other than English and Korean is not guaranteed. The developers warn that it may not provide reliable advice for professional fields like law, medicine, or finance. Factual inaccuracies and inappropriate expressions may still occur despite safety filtering.
How to use it
Weights are available under the MIT License. Users can run the model locally using tools like llama.cpp, LM Studio, or Ollama. It also supports deployment via vLLM with an OpenAI-compatible API. Friendli.AI offers hosted endpoints for immediate testing, though login is required after five interactions. Tutorials for setup are provided on GitHub.
Safety and testing
Researchers attempted to exclude profanity, discriminatory language, and bias from the training data. However, the documentation notes that the model may still generate inappropriate expressions or factual errors. It is not intended for high-stakes professional advice without human oversight.
Questions people ask about Mi:dm 2.0
What languages does Mi:dm 2.0 support?
The model is trained primarily on English and Korean data. The developers state that understanding and generation in other languages are not guaranteed.
Can I run this model on my own computer?
Yes. The 2.3 billion parameter version is optimized for on-device use. You can run it using llama.cpp, LM Studio, or Ollama on local hardware.
Is the model open source?
The weights are released under the MIT License, which is a permissive open license. You can download them from Hugging Face and use them freely within those terms.
Good reading and sources
There is no English Wikipedia article on Mi:dm 2.0 yet, so the facts on this page come from the sources listed above: huggingface.co/K-intelligence/Midm-2.0-Base-Instruct. Some are the subject's own pages, so its claims about itself are reported as its claims. EveryLLM wrote the text in its own words, with AI assistance, and compares it with the sources to make sure no passage is copied.
It may contain mistakes, and figures change over time. Nothing here is investment or buying advice. If something is wrong, tell us. Last refreshed 2026-11-14.