Veyra

Indonesian language models have historically struggled with a fundamental structural challenge: the language's rich system of affixes and word modifications gets fractured by tokenization approaches designed for English. Veyra, a 75-million-parameter language model, addresses this problem through an alternative architecture that treats Indonesian on its own terms.

The core innovation is the NMU framework—ninmeni meaning unit—which assigns a fixed ID to each character. Rather than breaking down inflected words into subword tokens, the model ingests complete words with all their affixes intact. This preserves the semantic structure that Indonesian speakers naturally recognize, training the model on authentic linguistic patterns instead of reconstructed approximations. The approach reflects a deliberate design philosophy: that language models should learn from a language's native roots, not translated or adapted paradigms.

The model was built entirely from scratch using this framework, with the 75M parameter size chosen as a conscious validation point. Developers can verify model capabilities through systematic evaluation rather than selective demonstration. The focus remains deliberately narrow: Indonesian language performance shapes corpus selection, evaluation priorities, and development direction.

Veyra shows unexpected capability beyond its primary scope. Despite no explicit English training, it generates grammatically sound English sentences—an artifact of how the NMU character-level encoding treats Latin characters universally across both languages. Developers treat this as an observable phenomenon rather than a marketed feature, documenting it while continuing to investigate the mechanism.

The product positions itself as a tool for builders working specifically with Indonesian language applications. By rejecting the usual transfer-learning approach that adapts English-trained models for other languages, it offers a model trained in the way Indonesian actually works. This appeals to developers seeking more authentic language understanding for Indonesian, teams building primarily for Indonesian-speaking users, and researchers interested in non-English-centric language model design.

The emphasis on documented evaluation and transparency about capabilities—including what remains unverified—indicates a research-first mindset that prioritizes credibility over marketing claims. No pricing information appears in available materials, suggesting this may be an open-source or research-stage project focused on validating the NMU framework before commercial deployment. That restraint itself signals maturity: the commitment to prove the approach works before scaling operations.