The model is explicitly identified as a dense decoder-only transformer following the Mistral-7B design pattern. Documentation confirms the use of standard refinements including Rotary Positional Embeddings (RoPE) and Grouped Query Attention (GQA). Technical specifications are detailed, including 32 layers, a hidden dimension of 4096, and a 4096-token context window. While the model is described as 'trained from scratch,' the reliance on the Mistral architecture is well-documented in the official repository and technical reports.

The training data is disclosed as a 349GB JSONL dataset (approx. 90 billion tokens) derived from 197 specialized Malaysian sources. Specific categories are named, including government documents (e.g., parliament transcripts), local news (Bernama, Star), and social media (Lowyat, Cari). The developers provide a list of scraped websites and reproduction notebooks for data collection. However, a precise percentage breakdown of the final training mixture (e.g., web vs. code vs. instructions) is not explicitly quantified in a single comprehensive table.

The model uses a custom Byte Pair Encoding (BPE) tokenizer with a 32,000-vocabulary size, specifically trained on a multilingual Malaysian corpus. It supports Malay, English, Mandarin, Tamil, and Jawi scripts. The tokenizer is publicly available on Hugging Face, and its alignment with the claimed language support is verifiable through the provided configuration files and training methodology documentation.

The model is clearly defined as a 7B parameter dense architecture. Unlike Mixture-of-Experts (MoE) models, all parameters are active during inference. The architectural configuration (32 layers, 4096 hidden size) is standard for this class and consistently reported across the official Hugging Face model card and GitHub repository.

Hardware specifications are partially disclosed; the developers mention using a Ray cluster with 5 nodes of 4x A100 80GB GPUs. While the hardware type is clear, the total GPU hours, training duration, and carbon footprint are not explicitly stated in the primary documentation. Some cost-efficiency claims (87% savings using AWS Trainium) are mentioned in press releases but lack the raw compute metrics required for a higher score.

The model has been evaluated on localized benchmarks like MalayMMLU and compared against ChatGPT 3.5. While some evaluation scripts are available in the repository, the exact prompts and few-shot examples used for the official reported scores are not fully documented in a centralized, reproducible format. Third-party verification is available through the MalayMMLU leaderboard, but the internal evaluation methodology remains partially opaque.

MaLLaM-7B consistently identifies itself as a Malaysian Large Language Model developed by Mesolitica. It does not exhibit identity confusion with other major models like GPT-4 or Llama. The versioning (e.g., v1.1, v2.5) is clearly tracked in the Hugging Face collections, and the model's limitations regarding its specialized regional focus are transparently discussed.

The model is released under the Apache 2.0 license, which is a standard, permissive open-source license. The terms for commercial use, modification, and redistribution are clear and consistent across the GitHub repository and Hugging Face model card. There are no conflicting proprietary restrictions mentioned in the official documentation.

VRAM requirements are documented for standard 16-bit (approx. 14-16GB) and 4-bit (approx. 4-5GB) inference. The model card provides sample code for loading with BitsAndBytes 4-bit quantization. While it lacks a detailed context-scaling memory table, the baseline requirements for consumer and enterprise hardware are well-defined.

The project uses version numbers (e.g., MaLLaM-7B, MaLLaM-v2), but it lacks a formal, detailed changelog or semantic versioning history that tracks specific weight updates or performance drift over time. Users must rely on separate Hugging Face model entries to track progress, which makes monitoring silent updates difficult.