Merdeka LLM is Agmo's answer to a specific problem: some of the most valuable data in Malaysia — legal, HR, financial, educational — is exactly the data that can't simply be sent to an overseas general-purpose model. Merdeka LLM is positioned as AI Sovereignty as a Service: developed by a local team, hosted in Malaysian data centres, trained on local data, so government and enterprise clients keep control of their data, hosting and model ownership, entirely locally.
An industry-specialist factory, not a general chatbot
Merdeka LLM was never built to be a general-purpose chatbot. Its job is helping different industries train their own specialist expert models — Malaysia Legal LLM, Malaysia HR LLM — fine-tuned on Malaysian legal, HR and industry-specific data, for tasks like contract review, compliance checks and policy document organisation.
Education: MerdekaLLM-Sasbadi-27b
MerdekaLLM-Sasbadi-27b is Merdeka LLM's education model, fine-tuned for Malaysian education in partnership with Sasbadi.
- Applications: personalised learning, curriculum development, and AI-driven tutoring platforms in both Malay and English.
- What it automates: personalised learning experiences for students across Malaysia, while aiding educators in curriculum planning and delivering digital education tools.
- Key benefit: tailored learning experiences, enhanced educational tools, and efficient education delivery.
Benchmark results: 84.9% on MalayMMLU
On MalayMMLU, the Malay-language knowledge benchmark, MerdekaLLM-Sasbadi-27b reaches 84.9% overall accuracy, the highest of the Malaysian-built models in our comparison. It is a 27B-parameter model that runs on a single GPU.
| Category | Accuracy |
|---|---|
| STEM | 87.49% |
| Language | 87.23% |
| Social science | 80.60% |
| Humanities | 87.31% |
| Others | 84.43% |
Small enough to host in a school
MerdekaLLM-Sasbadi-27b outscores ILMU GLM-5.1 (84.0%) on MalayMMLU with about 28× fewer parameters. At FP16 its weights need about 54 GB of memory, so it runs on one 80 GB GPU server. That is the difference between a server in a school and a data-centre cluster.
- Student data stays in school. Questions, answers and student work never leave the school's own server: sovereign by design, and simpler for PDPA compliance.
- A fraction of the hardware. One GPU server instead of a data-centre cluster, so a school or district can own its AI instead of renting it.
- No dependence on outside clouds. It runs on the school network, so lessons don't stop when the internet or an overseas provider does.
Internal evaluation, 26 August 2026, on the full MalayMMLU set (24,213 multiple-choice questions across 22 Malaysian school subjects; UM × YTL AI Labs, Findings of EMNLP 2024). Comparison data from the Pendakwah Teknologi MalayMMLU leaderboard, retrieved 7 October 2026. “Highest of the Malaysian-built models” refers to that comparison.
See the full model hub, size chart and leaderboard comparison at merdekallm.com/merdeka-model-llm.
Why sovereignty matters here
The real distinguishing feature isn't whether Merdeka LLM can write flowery prose — it's that in highly sensitive domains, it can ingest local data that can't go to an overseas cloud, and train a genuinely specialist model, within a local compliance environment, that actually understands local regulation, institutions and requirements. That kind of data is generally off-limits to global general-purpose models — and that's the core value sovereign AI delivers.
Merdeka LLM is a factory purpose-built for vertical specialists — working with experts in law, HR, education and finance to turn industry-specific knowledge into individual domain models under a sovereignty-first approach.