KT Mi:dm
KT's sovereign Korean-first LLM — MIT-licensed, enterprise-ready
Overview
KT Corporation's Korean-first LLM: quote-only enterprise tiers via Mi:dm Studio, plus the MIT-licensed Mi:dm 2.0 open weights (Mini 2.3B, Base 11.5B) on Hugging Face. The 32B Mi:dm K 2.5 Pro adds English, Japanese and Chinese for enterprise and public-sector deals.
Pricing
Pricing shown for reference only. These figures reflect RECATOOLS research as of 11 Jul 2026 and may be out of date or incomplete. This is not financial or purchasing advice — always confirm the current price on the provider’s official website before making any decision.
- Enterprise API access
- Quote-based pricing
- Higher throughput
- Quote-based pricing
- Advanced reasoning
- Quote-based pricing
- Highest-capacity model
- Dedicated account support
Use cases
What you can produce with KT Mi:dm
- A locally-running Korean-English bilingual chatbot built on Mi:dm 2.0 Base via Ollama or LM Studio
- A fine-tuned domain-specific Korean LLM using KT's SFTTrainer tutorials and trl library
- An OpenAI-compatible vLLM API server hosting Mi:dm 2.0 for internal enterprise use
- A Korean-language RAG system with reduced hallucinations using Document AI and Factguard AI techniques
- An on-device Korean AI assistant running Mi:dm 2.0 Mini on a mobile or edge device
- A function-calling agentic workflow with Mi:dm's custom tool parser and MCP server integration
- A benchmark comparison report of Mi:dm vs. competing Korean LLMs on KMMLU and HAERAE datasets
ASEAN Perspective
KT Mi:dm in Southeast Asia
Mi:dm is a Korea-first sovereign AI product with no confirmed ASEAN deployment or regional data-residency infrastructure as of mid-2026; its sovereign-cloud framing (data stays in Korean KT infrastructure) actually works against cross-border ASEAN adoption, where local data residency laws differ by country. The Mi:dm K 2.5 Pro's addition of Japanese and Chinese alongside Korean signals APAC awareness, but Southeast Asian languages (Malay, Indonesian, Thai, Vietnamese, Filipino) are absent from the training corpus. ASEAN enterprises that do engage KT would likely do so through its Microsoft or AWS partnership channels rather than directly, and Korean-language cultural grounding offers limited lift in SEA markets where English, Malay, and Bahasa Indonesia dominate enterprise AI use cases.
Mi:dm is KT's bet that a telecom-scale Korean LLM beats importing one. The MIT-licensed Mi:dm 2.0 Base (11.5B) was the first Korean commercial-grade open model over 10B parameters, and it holds its own against EXAONE-3.5 on KMMLU and HAE-RAE — a real result, not marketing. The enterprise side, Mi:dm K 2.5 Pro, now runs 32B parameters with 128K context and posted decent tool-use scores (87% on tau2-bench) at MWC26. Outside Korea the story thins out. Mi:dm Studio pricing is quote-only, docs lean Korean, and the ASEAN developer community has barely touched it. The 41B Pro variant is still private preview. For Korean enterprise — banking, public sector, telecom — this is a credible sovereign alternative to importing GPT or Gemini. For anyone else, it's a model to watch, not yet one to build on.
What people say
Eleven-and-a-half billion parameters, MIT license, first Korean model of that size to go fully open — that's the pitch for Mi:dm 2.0 Base, and Korean developers on Hugging Face have been receptive. The Mini variant (2.3B) runs on-device, and both outperform similarly-sized multilingual models on Korean commonsense benchmarks like HAE-RAE, though English reasoning still trails frontier models from OpenAI or Anthropic.
The enterprise side is a different animal. KT sells access through Mi:dm Studio in four tiers — Basic through Expert, the top end scaling toward 200B parameters — and none of it is priced in public. Buyers who want the 32B Mi:dm K 2.5 Pro, KT's newest release with 128K context and four-language support (Korean, English, Japanese, Chinese), go through direct sales. KT frames this as sovereign-cloud positioning for finance and public-sector clients wary of routing data through US hyperscalers, and Factguard, KT's hallucination-reduction layer, gets specific mention in its marketing.
Documentation is mostly Korean, and outside the domestic telecom and public-sector base there isn't much of a track record yet — no G2 or Capterra listing, no ASEAN case studies. The official GitHub repo (K-intelligence-Midm/Midm-2.0) has modest but real community activity. This is a company betting its size and government relationships turn into an AI moat, and for Korean-language enterprise work, that bet looks reasonable so far.
Summary of public user & expert reviews, compiled by RECATOOLS.
Notable facts
- Mi:dm's name comes from the Korean word 미듬 (mideum) meaning 'belief' or 'trust' — a deliberate cultural statement about AI reliability.
- KT became the first Korean company to open-source a commercial-grade LLM with over 11 billion parameters when it released Mi:dm 2.0 Base in July 2025.
- The original 2023 Mi:dm flagship model topped 200 billion parameters — larger than GPT-3 — making it one of the largest LLMs ever built by a telecom operator.
- Mi:dm K 2.5 Pro was built in collaboration with the 'K Data Alliance,' a consortium of Korean public institutions, academia, media organisations, and educational bodies.
Frequently asked questions
About this listing
This entry was compiled from publicly available data including KT Mi:dm's official website, press releases, documentation, and reputable third-party publications. RECATOOLS is not affiliated with KT Mi:dm unless explicitly stated.
Third-party AI tools update their pricing, features, availability, and policies frequently. Information here may be outdated by the time you read this — we make reasonable efforts to keep listings current, but cannot guarantee absolute accuracy.
For the latest details, please refer to KT Mi:dm directly →
Spotted something out of date? Suggest an update →
Alternatives to KT Mi:dm
More in LLMs & Chat
