Veyra

Unknown

• Updated Aug 12, 2026, 03:17 AM

Veyra is a series of Indonesian language models built on the native NINMENI framework, which uses the NMU (ninmeni meaning unit) paradigm where each character is assigned a fixed ID.

Use Cases ▼

Indonesian text generation

Natural language understanding for Indonesian

Morphological analysis and linguistic research

Building Indonesian chatbots and virtual assistants

Content creation and summarization in Indonesian

Educational tools for Indonesian language learning

What Problem It Solves ▼

Standard tokenization methods like BPE fracture Indonesian words, disrupting the language's rich system of prefixes, suffixes, and infixes, which alters meaning. NINMENI's NMU paradigm preserves word integrity by mapping each character to a fixed ID, allowing the model to learn morphological structures natively without prior statistical decomposition.

Key Features ▼

✓

Native support for Indonesian morphology via character-level IDs

✓

Transparent and auditable due to small model size

✓

Built from scratch without inheriting biases from other languages

✓

Layered knowledge training (language, facts, cultural context)

✓

Expanded character space in second generation for better coverage

Overview

Veyra is a series of Indonesian language models built on the native NINMENI framework, which uses the NMU (ninmeni meaning unit) paradigm where each character is assigned a fixed ID. This approach keeps words intact with their affixes, addressing the morphological complexity of Indonesian that typical subword tokenization fragments. The first generation, Veyra 75M, demonstrated the framework's viability, and the second generation, Veyra Nyra 92M, is trained from scratch on a refined framework with expanded character space. The framework is designed specifically for Indonesian, not adapted from multilingual models, and the models are intentionally compact to enable rigorous validation and auditing of their internal representations.

Pros & Cons

Pros

Native support for Indonesian morphology via character-level IDs

Transparent and auditable due to small model size

Built from scratch without inheriting biases from other languages

Layered knowledge training (language, facts, cultural context)

Expanded character space in second generation for better coverage

Cons

−

Small model size may limit performance on complex tasks

−

Currently limited to Indonesian language only

−

Second generation still in training, not yet finalized

−

May lack broad world knowledge compared to larger models

−

Niche focus may not suit multilingual applications

Similar Tools

IndoBERT IndoNLG ByT5 mBART-50 XLM-R GPT-3 Indonesian fine-tunes