Open Foundation Model
for Korea and Beyond

Lotus is a long-term research initiative to build an open, sovereign, and reproducible foundation model ecosystem based on open data, open models, open training code, and open evaluation.

Building independent AI infrastructure

Foundation models are becoming core infrastructure for national competitiveness, industrial productivity, and future AI services. Lotus aims to reduce dependency on external AI platforms and establish a sustainable open AI ecosystem for Korean language, knowledge, and industry-specific intelligence.

Open FoundationOpen models, datasets, code, checkpoints, and results.
ReproducibilityDocumented training settings and repeatable pipelines.
Incremental GrowthStep-by-step scaling from 1B to 123B parameters.
Community DrivenBuilt with researchers, developers, and partners.

From 1B to 123B parameters

Lotus follows an incremental scaling strategy. Each phase expands model capacity, training tokens, Korean language capability, reasoning quality, coding performance, and domain applicability.

Model Scale ProgressionParameter size by development phase
1B
3B
7B
13B
30B
70B
123B
Phase 1Lotus-1B
20B ~ 40B
2026.08
Foundation pipeline validation
Phase 2Lotus-3B
50B ~ 100B
2026.10
Practical baseline and code generation
Phase 3Lotus-7B
100B ~ 300B
2026.12
Korean language performance and industry use
Phase 4Lotus-13B
300B ~ 500B
2027.03
Domain expertise and long-context reasoning
Phase 5Lotus-30B
500B ~ 1T
2027.06
Enterprise productivity optimization
Phase 6Lotus-70B
1T ~ 2T
2027.09
High-performance general-purpose LLM
Phase 7Lotus-123B
2T ~ 5T
2027.12
Sovereign LLM completion

High-quality corpus and reproducible data pipeline

Lotus is designed around a large-scale Korean and English training corpus. Raw data is processed through collection, cleaning, validation, deduplication, quality scoring, and publishable dataset packaging.

1TB+Korean training corpus secured
75%English data for reasoning, math, and code
25%Korean data for language specialization
01Collect

Web, Wikipedia, Wikisource, news, literature, documents

02Clean

HTML removal, spam filtering, broken document removal

03Validate

Language checks, length checks, deduplication, scoring

04Publish

JSONL, Parquet, metadata, Hugging Face dataset release

Llama-based decoder-only transformer

Lotus adopts a modern decoder-only transformer architecture to reduce training risk and maximize compatibility with the open source LLM ecosystem. The stack is optimized for scalable pretraining, efficient inference, and reproducible release.

LotusFoundation Model
Decoder-only TransformerRMSNormSwiGLURoPEGrouped Query AttentionFlashAttention 2BF16 Mixed PrecisionPyTorch FSDPSentencePiece BPEvLLM

From base model to domain expert models

Lotus evolves through four model layers: Base, Instruct, Chat, and Domain Expert. This structure enables the project to move from general language modeling to practical Korean enterprise use cases.

Phase 1Base Model

Next-token prediction

Phase 2Instruction Model

Supervised fine-tuning

Phase 3Chat Model

DPO / ORPO / SimPO

Phase 4Domain Expert

Domain-adaptive continued pretraining

Specialized intelligence for real-world industries

On top of the foundation model, Lotus will support domain-specific expert models for legal, public, software, IT, semiconductor, and financial applications.

Lotus-LawLegal documents, precedents, administrative texts
Lotus-PublicPublic administration, civil service, government AI
Lotus-CodeSoftware development, code generation, code review
Lotus-ITIT industry news, technical articles, ICT policy
Lotus-SemiconductorSemiconductor process, device, and design documents
Lotus-FinanceFinance, economy, investment, and accounting documents

Open models, open code, open evaluation

Lotus aims to release datasets, tokenizer, training code, preprocessing scripts, SFT and preference alignment code, checkpoints, quantized models, and benchmark results through open platforms such as Hugging Face and GitHub.

Datasets
Tokenizer
Training Code
Evaluation Code
Checkpoints
Benchmark Results
Building Korea's Open Foundation Model