Mesolitica
Malaysia-based AI lab building MaLLaM, open-weight LLMs pre-trained from scratch on Malay text
Mesolitica is an open-source Malay-language LLM lab best for Malaysian teams that need Bahasa Malaysia AI models hosted on their own infrastructure rather than a foreign API. Its SEA edge is MaLLaM, trained from scratch on a 90-billion-token Malay corpus instead of translated from English, published as free open weights on Hugging Face in 1.1B, 3B, and 5B sizes. Caveat: there is no managed API, so teams must provision and operate their own GPU or Ollama infrastructure to run the models in production.
- ✓Trained from scratch on Malay text rather than fine-tuned from an English base model
- ✓Free open-weight release removes per-token API costs
- ✓Self-hosting keeps inference data inside Malaysia for regulated use cases
- ✓Multiple parameter sizes (1.1B/3B/5B) let teams match model size to available hardware
- ×No managed API or hosted endpoint; deployment requires operator-provisioned GPU infrastructure
- ×5B variant needs production-grade GPU hardware, raising entry cost versus a hosted API
- ×Smaller ecosystem and documentation base than global LLM providers
- ×Coverage limited to Malay and Malaysian English rather than broader SEA languages
About Mesolitica
Mesolitica is a Malaysia-based AI lab that develops MaLLaM, a family of large language models pre-trained from scratch on a Malay-language corpus rather than fine-tuned from an English base model. The models ship as 1.1B, 3B, and 5B parameter checkpoints with instruction-tuned chat variants, alongside separate Malaysian speech-to-text and text-to-speech model collections, all distributed as free open weights on Hugging Face for self-hosted deployment in Malaysia, Singapore, and Indonesia.
Key Features
Best For
Southeast Asia Fit
MaLLaM is pre-trained from scratch on Malay text rather than fine-tuned from an English base model, and ships as open weights that can be self-hosted inside Malaysia for data residency, unlike most foundation models that route inference through US-based APIs.
- Malaysia
- SEA
Integrations
- Other
- Hugging Face Transformers
- Ollama
- AWS
This profile was researched and written with AI from Mesolitica's public pages, and AI assistants quote pages on this site. Something wrong or out of date? Email [email protected] with this page and a public source. Corrections are free, for every listing.
How listings and corrections work →"Verified" means read on the vendor's own published pages on the date shown. We do not run hands-on tests, and a statement a vendor confirmed to us directly is labelled as vendor-confirmed. Software-listing.com is independent and may earn affiliate commissions from some links.
Related Analysis & Guides
AI Coding Assistants for SEA Developers in 2026: Cursor, Copilot, and What's Actually Worth Paying For
AI for KYC and Document Processing: How SEA Fintechs Build Onboarding in 2026
Clinic Software Doesn't Cross Borders: Why a Thai Interface Isn't Thai Localisation
The questions operators actually ask.
Is Mesolitica's MaLLaM free to use?
Yes. MaLLaM model weights are released as open-source on Hugging Face at no cost; the only expense is the compute needed to self-host inference.
Does MaLLaM require sending data to Mesolitica's servers?
No. MaLLaM runs on infrastructure the operator controls, such as a workstation, on-premise GPU, or cloud instance, with no inference calls routed to Mesolitica or any third party.
What languages does MaLLaM support?
MaLLaM is trained primarily on Malay-language (Bahasa Malaysia) and Malaysian English text, rather than being a general multilingual model fine-tuned after the fact.