Our Model Hub
Merdeka LLM is a family of domain-specialised models, each fine-tuned with a partner who knows the field. Here are its real-world applications across Malaysia’s priority sectors. Hold quality data in your field? Partner with us on the next model.
🎓 Education: MerdekaLLM-Sasbadi-27b
Fine-tuned for Malaysian education in partnership with Sasbadi.
Applications
Personalised learning, curriculum development, and AI-driven tutoring platforms in both Malay and English.
Automations
Merdeka LLM can create 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.
MalayMMLU benchmark
84.9%
overall accuracy
#1 Malaysian model
27B · runs on one GPU
| Category | Bar | Accuracy |
|---|---|---|
| STEM | 87.49% | |
| Language | 87.23% | |
| Social science | 80.60% | |
| Humanities | 87.31% | |
| Others | 84.43% |
Internal evaluation, 26 August 2026, using MalayMMLU (UM × YTL AI Labs, Findings of EMNLP 2024), the full set of 24,213 multiple-choice questions across 22 Malaysian school subjects. Overall accuracy is across all questions, so it weights each category by its size. See how it compares · Why size matters.
⚖️ Legal
New version coming soonApplications
Contract review automation, legal research assistance, document summarisation, and legal compliance checks.
Automations
Merdeka LLM can automate tedious legal tasks such as document drafting, contract reviews, and legal research, ensuring compliance with Malaysian laws and regulations.
Key Benefit
Increased legal department productivity, reduced manual workload, and enhanced accuracy in legal operations.
Become our legal data partner
We’re building the next Merdeka legal model with a partner who knows the field: law firms, legal publishers and professional bodies with quality Malaysian legal content.
👥 Human Resources (HR)
New version coming soonApplications
Automated resume screening, employee onboarding, compliance training, and performance evaluations.
Automations
HR teams can leverage Merdeka LLM to automate key tasks such as resume filtering, employee evaluations, and regulatory compliance training, streamlining recruitment and management.
Key Benefit
Efficient hiring processes, improved employee engagement, and better overall HR operations with reduced human bias.
Become our HR data partner
We’re building the next Merdeka HR model with a partner who knows the field: HR consultancies, payroll and HRMS providers, and training bodies with quality Malaysian HR content.
💰 Finance
Coming soonApplications
Tax advisory, tax planning, and automated customer service solutions.
Automations
Leverage Merdeka LLM to streamline customer interactions, enhance security, and provide predictive financial insights, all while ensuring compliance with local regulations.
Key Benefit
Optimised operations, secure financial analysis, and enhanced customer experiences.
Become our finance data partner
We’re building the next Merdeka finance model with a partner who knows the field: banks, tax advisory firms and financial publishers with quality Malaysian finance content.
Small enough to host in a school
At 27B parameters, MerdekaLLM-Sasbadi-27b runs on a single GPU server, so a school can host it on its own premises.
- MerdekaLLM-Sasbadi-27b
- ILMU GLM-5.1
- Other models
Chart data and sources
| Model | Parameters | MalayMMLU | Size source |
|---|---|---|---|
| Kimi K3 | 2.8T | 87.5 | Model card |
| DeepSeek V4 Pro 0813 | 1.6T | 85.5 | Model card |
| Qwen 3.5 397B | 397B | 85.4 | Model card |
| MerdekaLLM-Sasbadi-27b | 27B | 84.9 | Agmo Group |
| ILMU GLM-5.1 | 744B | 84.0 | Model card |
| DeepSeek V4 Pro | 1.6T | 83.8 | Model card |
| MiniMax M3 | 428B | 82.8 | Model card |
| GLM-5.2 | 744B | 81.8 | Model card |
| Mistral Large 3 | 675B | 81.5 | Model card |
| Kimi K2.6 | 1T | 80.7 | Model card |
| Mistral Large 2512 | 675B | 80.4 | Model card |
| MiniMax M3 | 428B | 79.6 | Model card |
| Llama 4 Maverick | 400B | 78.1 | Model card |
| GLM-5.3 | 744B | 71.6 | Model card |
ILMU GLM-5.1 is fine-tuned from Z.ai’s GLM-5.1, so it has the same parameter count. Accuracy from the Pendakwah Teknologi leaderboard, except MerdekaLLM-Sasbadi-27b (internal evaluation). Total parameters; mixture-of-experts models use only part of them per token but must hold all of them in memory.
| Measure | MerdekaLLM-Sasbadi-27b MY | ILMU GLM-5.1 MY |
|---|---|---|
| MalayMMLU accuracy | 84.9% | 84.0% |
| Parameters | 27B | 744B (mixture-of-experts) |
| Memory for the weights | ~54 GB | ~1.5 TB |
| Hardware to host it | One 80 GB GPU | ~19 × 80 GB GPUs |
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.
Modest hardware
One GPU server is enough, 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.
Memory and GPU counts are for the model weights alone. Scores from the MalayMMLU comparison below.
How we sized it
- Memory is for the model weights alone; running a model also needs some extra memory for the conversation itself.
- ILMU GLM-5.1 is fine-tuned from Z.ai’s GLM-5.1, and fine-tuning does not change a model’s parameter count (744B per Z.ai’s GLM-5 repository).
- A mixture-of-experts model uses only part of its parameters for each token, but all of them must be held in memory to serve it.
How MerdekaLLM-Sasbadi-27b compares
Accuracy (%) on MalayMMLU, the Malay-language knowledge benchmark. At 84.9%, MerdekaLLM-Sasbadi-27b scores highest of the Malaysian-built models (MY) in this comparison and sits among frontier models from Google, OpenAI and Anthropic.
| Model | Parameters | Overall | STEM | Language | Social science | Humanities | Others |
|---|---|---|---|---|---|---|---|
| Gemini 3.1 ProGoogle | Undisclosed | 91.5 | 91 | 96 | 89 | 93 | 89 |
| GPT-5.5OpenAI | Undisclosed | 89.3 | 89 | 93 | 85 | 90 | 89 |
| Claude Opus 4.8Anthropic | Undisclosed | 88.7 | 86 | 90 | 86 | 92 | 89 |
| MerdekaLLM-Sasbadi-27b MY OursAgmo Group × Sasbadi | 27B | 84.9 | 87 | 87 | 81 | 87 | 84 |
| ILMU GLM-5.1 MYYTL AI Labs | 744B | 84.0 | 85 | 86 | 80 | 85 | 84 |
| ILMU v3.1 MYYTL AI Labs | Undisclosed | 83.1 | 84 | 87 | 80 | 84 | 81 |
| ILMU Vision v1.3 MYYTL AI Labs | Undisclosed | 74.9 | 79 | 75 | 73 | 74 | 74 |
| ILMU Mini v3.3 MYYTL AI Labs | Undisclosed | 71.8 | 74 | 72 | 72 | 71 | 70 |
Top 3 leaderboard models and every Malaysian model shown. Kimi K3, Qwen 3.7 Max, DeepSeek V4 Pro 0813 and Qwen 3.5 397B also score above ours (85.4 to 87.5); the rest are on the full leaderboard.
Source: Pendakwah Teknologi, 7 October 2026. Ours is an internal evaluation.
How we compared
- MerdekaLLM-Sasbadi-27b was scored the same way as the leaderboard (generate-and-parse) on the full 24,213-question MalayMMLU set; the leaderboard samples 1,100 of these.
- It is not a leaderboard entry. Its category scores are rounded to match the leaderboard.
- “#1 Malaysian model” compares the Malaysian-built models in this table.
- MalayMMLU by Poh et al., UM × YTL AI Labs (Findings of EMNLP 2024).
Let's Shape the Future Together
Join us in building a brighter tomorrow, reach out anytime.
Contact details
Level 38, MYEG Tower,
Empire City Damansara,
Jalan PJU 8, Damansara Perdana,
47820 Petaling Jaya, Selangor, Malaysia.