MaLLaM
Mesolitica's from-scratch Malaysian models, free to download but unlicensed
Overview
MaLLaM (Malaysia Large Language Model) is a family of 1.1B, 3B and 5B base models from Malaysian startup Mesolitica, pre-trained from scratch on about 90 billion tokens of Malaysian text with the Mistral architecture, plus instruction-tuned variants. The weights are free on Hugging Face but declare no licence. AWS said in 2024 that a larger MaLLaM had been trained on close to 200 billion tokens; it has not been released.
Pricing
Pricing shown for reference only. These figures reflect RECATOOLS research as of 19 Oct 2026 and may be out of date or incomplete.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.
What you can produce with MaLLaM
- Download base models in 1.1B, 3B and 5B sizes, pre-trained from scratch on about 90 billion Malaysian tokens
- Try instruction-tuned 1.1B, 3B and 5B variants with a 20,480-token context
- Load them as standard Mistral-architecture models in Hugging Face transformers, including 4-bit quantisation
- Self-host Malay text generation inside Malaysia rather than sending text to a foreign API
- Fine-tune a small Malay base model on your own data for research
- Read the arXiv technical report and training logs behind the 2023 models
ASEAN Perspective
MaLLaM in Southeast Asia
MaLLaM is the most substantial open Malay model available, and Mesolitica trained it from scratch on roughly 90 billion tokens of Malaysian text rather than fine-tuning an English base. In practice, this means it handles not just formal Bahasa Malaysia but also the slang, dialect, and code-mixing people actually use—the registers where models trained on formal corpora fail. It ships at 1.1B, 3B and 5B, small enough to run on modest hardware, which matters for Malaysian institutions without GPU budgets. Weights are on Hugging Face; the reputation lives in the Malaysian AI community rather than on any review site.
MaLLaM is a genuine from-scratch Malaysian model and a useful research artefact: base checkpoints from 1.1B to 5B parameters and instruction-tuned versions with a 20,480-token context, built on Malaysian text rather than translated through English.
For anything beyond research, the lack of any declared licence is disqualifying until Mesolitica adds one, and the line has not been updated since 2024. The larger AWS-trained model was never published. Malaysian teams wanting production Malay today should start with a maintained, licensed multilingual model.
What we found
MaLLaM (Malaysia Large Language Model) is a set of language models from Mesolitica, a Malaysian AI startup, pre-trained from scratch with the Mistral architecture on about 90 billion tokens of Malaysian text, roughly 349 GB, rather than adapted from an English base. The checkpoints appeared on Hugging Face in November 2023 in 1.1B, 3B and 5B sizes with a 4,096-token context, followed in December by instruction-tuned versions with a 20,480-token context. A technical report on arXiv in January 2024 compared the instruction-tuned models with ChatGPT-3.5 and Malaysian Mistral and called them competitive, on Mesolitica's own evaluation.
The public story then moved to AWS. In a December 2024 media alert, AWS said Mesolitica had trained MaLLaM on 197 datasets totalling close to 200 billion tokens, handling Malay, slang and 16 other regional languages, and had cut compute costs by 87% while raising throughput 5.5 times on Trainium and Inferentia chips. Those are vendor figures. The alert quoted a health-tech company using MaLLaM for medical note-taking and said the Malaysian government was exploring it. That larger model has not been published; the open checkpoints are still the 90-billion-token ones.
Activity has since stopped. The MaLLaM repositories were last modified in October 2024, and Mesolitica's Hugging Face organisation, 278 models, was last updated in September 2025 with speech and translation releases. Its Malay LLM leaderboard is no longer public and its benchmark repository last changed in April 2024. Mesolitica's website is a single line linking to Hugging Face, with no pricing, API or hosted endpoint.
The licence is the practical problem. None of the eight MaLLaM repositories declares one, so there is no grant to rely on, which matters more to a commercial adopter than a restrictive licence would. The models are also small and need fine-tuning, and the instruct versions call themselves a quick demonstration with minimal moderation. They suit research on Malaysian Malay, not production.
Compiled by RECATOOLS from public sources: vendor documentation and pricing, code repositories and published reporting.
Frequently asked questions
About this listing
This entry was compiled from publicly available data including MaLLaM's official website, press releases, documentation, and reputable third-party publications. RECATOOLS is not affiliated with MaLLaM 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 MaLLaM directly →
Spotted something out of date? Suggest an update →
Alternatives to MaLLaM
More in LLMs & Chat
