PLANNEDRAYA MODEL
RAYA LOCAL / GENERAL
Indonesian-first general language profile for knowledge work, content, service, and structured generation.
GENERATION · STRUCTURED OUTPUTS01 / RAYA MODELSRAYA.COM.AI
Open models, localized for Indonesia.
RAYA selects capable open models and adapts them into local model profiles for Indonesian language, knowledge, culture, and commercial work.
POSITION
We do not localize at the prompt layer. Language, knowledge, behavior, and evaluation are addressed across the model lifecycle.
LOCALIZATION STACKRAYA / MODEL LAB
L04RAYA MODEL PROFILECommercial behavior · safety · tools
L03POST-TRAININGInstructions · preferences · domain tasks
L02CONTINUED PRE-TRAININGIndonesian language · local knowledge
L01SELECTED OPEN MODELCapability · license · architecture · fit
MODEL PORTFOLIO / DIRECTION
Planned RAYA profiles
Product names and availability remain subject to model evaluation and launch decisions. No performance specifications are claimed here.
PLANNEDRAYA MODEL
RAYA LOCAL / REASON
A reasoning-oriented profile for analysis, multi-step business tasks, and complex document workflows.
REASONING · ANALYSISPLANNEDRAYA MODEL
RAYA LOCAL / COMPACT
A smaller profile intended for responsive, repeatable commercial tasks where efficiency matters.
CLASSIFICATION · EXTRACTIONPLANNEDRAYA MODEL
RAYA EMBED / MULTILINGUAL
Embedding profile for Indonesian and multilingual retrieval, semantic search, and enterprise knowledge.
EMBEDDINGS · RETRIEVALPRODUCT READINESS
From candidate to RAYA release
Confirm commercial rights, obligations, model design, serving fit, and migration constraints.
Build governed language and knowledge mixtures, then document data scope and training decisions.
Post-train for instructions, tools, structured work, refusal behavior, and intended commercial tasks.
Evaluate quality, risk, reliability, and regressions before assigning a stable RAYA model version.
PUBLIC STATUSMODEL PORTFOLIO / PRE-LAUNCH
Model names, identifiers, specifications, availability, and underlying open-model selections remain subject to evaluation.
01 / APPROACH
How it works
Open model selection
Models are assessed by capability, license, architecture, language coverage, and target workload.
Continued pre-training
Governed Indonesian and domain corpora deepen language fluency and local knowledge.
Post-training
Instruction tuning, preference optimization, and safety alignment shape useful business behavior.
Local evaluation
Models are tested against Indonesian tasks, enterprise scenarios, and defined risk profiles.
02 / SCOPE
What it includes
Language
Bahasa Indonesia, code-switching patterns, regional language context
→Knowledge
Indonesian institutions, markets, terminology, and domain corpora
→Behavior
Commercial instructions, structured outputs, tools, and safety policies
→Delivery
Shared Token API, dedicated model profile, or private deployment
→03 / QUESTIONS
Frequently asked
Does RAYA train models from scratch?
RAYA’s core approach is to select capable open models and invest in deep Indonesian localization rather than recreate every base capability from zero.
What makes a model local?
Localization spans language data, local knowledge, commercial instructions, safety behavior, tools, and evaluation—not simply a system prompt.
Can a model be adapted to one industry?
Yes. A RAYA local profile can be extended with governed domain data and task-specific post-training where the business case supports it.