MaLLaM (Mesolitica)¶

FreeProfessional

Vendor: mesolitica.com | Type: Non-detector | Cost: Free

TL;DR

Mesolitica's open-source Malay-specific foundation model family, pre-trained from scratch on 90 billion Malaysian- context tokens, distributed through Hugging Face under the Mistral architecture in 1.1B and 5B parameter sizes. The toolkit's escalation-option in 2B.3 for Malaysian work where Malay-first verification phrasing matters more than broader regional coverage.

What it does¶

MaLLaM is a family of open-weight Malay-language foundation models pre-trained from scratch on a 90-billion-token corpus of Malaysian text. The current Hugging Face checkpoints are mallam-1.1B-4096 (1.1 billion parameters, 4096 context length, most recently updated 2 March 2026) and mallam-5B-4096 (5 billion parameters, updated 28 June 2025), with instruction-tuned variants for chat-style use. The deliberate design choice that distinguishes MaLLaM from broader regional LLMs is "pre-trained from scratch" rather than fine-tuned on Malay data: the model's representations are tuned to Malaysian linguistic idiosyncrasies from the ground up rather than inheriting English-language priors with Malay overlay.

For a fact-check workflow on Malaysian content, the operational role is Malay-first drafting and reasoning where phrasing, register, and Malaysian-context fluency matter to the editorial output. Where SEA-LION is the broader regional default covering 11 SEA languages including Malay, MaLLaM is the deeper Malay-specific option. The two are not substitutes; MaLLaM enters the toolkit as an escalation-option for cases where the Malay-language quality of the output is itself the selection criterion.

When to use it¶

A Malaysian newsroom or research lab needs Malay-language drafting, summarisation, or reasoning where Malay-first fluency matters more than the multi-language coverage SEA-LION provides.
An institutional partner is building a Malaysia-specific fact-check pipeline that requires a deeply Malay foundation model and acceptable infrastructure overhead for self-host.
A research project on Malay-language information environments needs a documented Malaysian-context-trained foundation model with auditable training-data provenance.
A regional partner pairing a Malaysian deployment with the broader-regional SEA-LION primary needs the Malay-specific deep option for the cases where SEA-LION's broader-regional default does not produce acceptable Malay output.

Limitations¶

Limitations

Local. MaLLaM is built on Malaysian contexts; it is not designed for cross-country regional deployment beyond Malay-language work. For Bahasa Indonesia (close but distinct linguistic environment), Filipino, Thai, Tamil, Sinhala, or Lao content the operational paths are SEA-LION (regional) and the country-specific 2B.2 entries.
LLM hallucination caveat is binding. Like every LLM, MaLLaM can produce confident-sounding output that is factually wrong. Output is suggestive at best; every claim must be verified against an independent source before publication.
Model size is modest by 2026 LLM standards (1.1B and 5B checkpoints). Performance on complex reasoning tasks may be more limited than the 27B-parameter SEA-LION v4 checkpoint; institutional partners should validate against their specific use case before broad adoption.
Sinhala, Tamil, Lao, Thai, Filipino are not covered; MaLLaM is single-language by design.
Operational deployment requires Hugging Face transformers infrastructure and developer integration; not a press-button tool for front-line fact-checkers without engineering support.

Privacy and threat model¶

The deployment posture is the privacy-defining choice, the same as for SEA-LION. Run MaLLaM on infrastructure the operator controls (workstation, on-premise GPU, or operator- provisioned cloud instance) and the data flow is fully under the operator's control. No input passes to Mesolitica or any third party at inference time; the open weights are downloaded once and run locally.

For Malaysian fact-check work in the regional research operating environment (CMA Section 233, Online Safety Act 2025 in force from 1 January 2026, 3R speech laws, the Malaysiakini January 2025 device-seizure precedent), the locally-deployable LLM path matters operationally. Institutional partners running MaLLaM on their own infrastructure can integrate Malay-language drafting into fact-check workflows without inheriting third-party cloud LLM threat-model implications.

For Malaysian source handling, the same general discipline applies: classify the input, route source-identifying material through human-mediated workflows rather than LLM-mediated ones, and prefer the institutional civil-society paths (Meedan Check, regional partner relays) over public- facing LLM interfaces for sensitive cases.

Country and platform applicability¶

Decision 7 framing applies: MaLLaM is an NLP-class tool, and content-language coverage is the operational selection criterion. Malay is the documented operational language; the malay-only-escalation-option flag is the explicit acknowledgement that this tool covers Malay only and enters the shortlist as an escalation-option rather than a primary.

Indonesia: not applicable. Bahasa Indonesia is linguistically adjacent to Malay but the corpus and tuning are Malaysian-context-specific; Indonesian content routes through Yudistira (MAFINDO Bahasa-specific) and SEA-LION (regional Bahasa support).
Laos: not applicable. Lao is not covered; the Lao path in 2B.3 is SEA-LION (which covers Lao explicitly per its v4 release documentation).
Malaysia: primary deployment context. Pairs with Sebenarnya AIFA at 2B.1 (the Malaysian institutional reference, with the documented government- operated independence caveat) and with SEA-LION at 2B.3 (broader-regional default; MaLLaM is the Malaysia-specific escalation-option for Malay-first phrasing). Documented use is Malaysian NLP.
Philippines: not applicable; Filipino / Tagalog is not covered.
Sri Lanka: not applicable; Sinhala and Tamil are not covered.
Thailand: not applicable; Thai is not covered.

Platform applicability: not platform-specific; MaLLaM is a foundation model. Its outputs feed whichever editorial workflow the operator runs at the front-end.

How to access¶

Open-source weights on Hugging Face under the Mesolitica organisation. The current checkpoints are mallam-1.1B-4096 (updated 2 March 2026) and mallam-5B-4096 (updated 28 June 2025), with instruction-tuned variants (mallam-1.1b-20k-instructions, mallam-1.1b-20k-instructions-v2) for chat-style use. Mesolitica also publishes a Malay LLM Leaderboard space on Hugging Face for cross-model comparison within the Malay-language category.

Deployment is via Hugging Face transformers, Ollama, or other standard LLM serving infrastructure. For institutional partners, the GitHub MaLLaM wiki under the malaysia-ai/malaya project documents architecture and training details; the Mesolitica organisation page on Hugging Face is the canonical distribution channel.

Cost (current as of 2026-05)¶

Free under the open-source licence. Operational cost is the operator's compute infrastructure: workstation-class hardware for the 1.1B variant, more substantial GPU for the 5B variant in production, plus developer time on integration. The 1.1B size profile is operationally lighter than SEA-LION's 27B multimodal variant at the cost of capability; the trade-off is Malay-specific tuning at smaller model size versus broader regional coverage at larger scale.

Verify current Hugging Face availability and any change in licence terms via the Mesolitica organisation page on Hugging Face directly before integrating.

Quickstart¶

(Quickstart for an institutional or developer deployment; front-line fact-checkers reach MaLLaM through their institutional partner's deployment.)

Confirm the use case is Malay-language editorial drafting or reasoning where Malay-first phrasing matters; for broader-regional or non-Malay cases route to SEA-LION.
Select the MaLLaM variant: mallam-1.1B-4096 for lighter deployments, mallam-5B-4096 for cases where the additional parameter count justifies the infrastructure overhead, or one of the instruction-tuned checkpoints for chat-style prompting.
Pull the weights from Hugging Face into the operator's serving infrastructure (Hugging Face transformers, Ollama, vLLM, or other supported runners).
Build prompt templates for the specific Malay-language editorial tasks: summarisation of Malaysian press, drafting of Malay-language clarification questions for a source, reasoning over Malaysian-context terminology that broader regional models stumble on.
Validate output on local Malaysian content before integrating into the live editorial workflow; LLM hallucination is binding and Malaysian-context-specific validation surfaces where the smaller model produces confident fabrications.
Pair the MaLLaM drafting output with an explicit human verification step, the same as for any LLM-mediated tool in 2B.3. Every model-drafted claim is verified against an independent source before publication.
For comparative deployment alongside SEA-LION (institutional workflows running both models), build the routing logic so Malay-first cases default to MaLLaM and broader-regional or non-Malay cases default to SEA-LION, with the operator reviewing output quality before settling the routing.

In the toolkit's workflow¶

Source-history pillar non-detector tool per Architectural Anchor 1. Sits in 2B.3 LLMs in fact-check workflows as the escalation-option for Malay-specific work alongside the institutional SEA-LION primary, the cloud-LLM Pinpoint alternative (cross-cell from 2A.1), and the mobile lay-user AI Fact Checker App alternative. MaLLaM's place in the cell is depth rather than breadth: when Malay-language quality matters more than coverage, MaLLaM escalates from SEA-LION.

Standard combinations:

With SEA-LION at 2B.3 – broader-regional default; MaLLaM is the Malay-specific deep option. The two cards together illustrate the depth-vs-breadth axis of SEA-region LLM choice.
With Sebenarnya AIFA at 2B.1 – Malaysian institutional reference (with the documented independence caveat). Civil-society and newsroom partners running Check-coalition workflows on Malaysian content can deploy MaLLaM as the LLM-mediated drafting layer alongside observation of AIFA outputs as documentary evidence about the Malaysian information environment.
With Sinar Project iMAP at 2C.1 – independent network-monitoring layer for Malaysian content, including detection of MCMC blocking events on competing outlets per regional case documentation. iMAP and MaLLaM are complementary parts of the civil-society-operated alternative to the government-operated AIFA.
With Meedan Alegre at 2B.2 – multilingual matching engine including Malay; pairs with MaLLaM for Malay-content tipline workflows where Alegre handles duplicate detection and MaLLaM handles editorial drafting.
With AI Fact Checker App at 2B.3 – Malaysian app-store localisation at the lay-mobile layer; MaLLaM is the institutional self-host alternative for cases where the LLM-mediated drafting output is part of a published workflow rather than lay triage.

Decision-tree references: T5 escalation references MaLLaM at T5.9 (text professional) for Malay-language LLM-mediated editorial drafting where Malay-first phrasing matters; broader-regional or cross-language cases default to SEA-LION.

This card carries no detector signal class: MaLLaM produces non-detector source-history signals (drafted summaries, candidate questions, translation-assist outputs supporting human verification). Per Anchor 2 LLM output is suggestive at best; every claim must be verified against an independent source before publication, and the verification step is mandatory rather than discretionary.

Override notes¶

Override notes

Malay-only escalation-option (malay-only-escalation-option): MaLLaM enters the 2B.3 shortlist as the Malay-specific escalation-option rather than a regional primary. The card surfaces this in the TL;DR, the Limitations admonition, and the Country and platform applicability section's Decision 7 framing. Cross-country deployment beyond Malay-language work is not the design intent and is not the operational use the toolkit recommends.

Sources¶

Mesolitica MaLLaM Hugging Face collection (mallam-1.1B-4096 updated 2 March 2026; mallam-5B-4096 updated 28 June 2025; instruction-tuned variants; pre-trained from scratch on 90B Malaysian-context tokens; Mistral architecture; 4096 context length)
MaLLaM arXiv paper (architecture and training details)