Open models. Efficient intelligence. Built from the ground up.

SOVYN is an independent AI research project focused on building
efficient language and multimodal models that developers can actually run.

# SOVYN 2.3 ### Current Generation

One model family.
Different levels of compute, speed and capability.

SOVYN 2.3 Fast Project Meridian The primary development model of the SOVYN 2.3 generation. ~1B PARAMETERS   8K CONTEXT

Architecture Decoder-only Transformer 28 Transformer Blocks ├── RMSNorm ├── RoPE ├── GQA ├── QK-Norm ├── SwiGLU └── Hybrid Attention ├── Local × 21 └── Global × 7 48K Tokenizer 8K Context Window

# Model Family

Designed as one generation with multiple deployment targets.

Mini Small footprint. Local-first deployment. PLANNED

Fast Fast inference. Efficient compute. IN DEVELOPMENT

Pro Capability and efficiency combined. PLANNED

Max The largest model in the generation. PLANNED

# Generations

01 SOVYN 1 First Generation Compact language models and foundational experiments. 85M

02 SOVYN 2 Second Generation Expanded model scale and training research. 300M

2.3 SOVYN 2.3 Current Generation Architecture, efficiency and multimodal research. MINI   FAST   PRO   MAX

# Research

SOVYN research spans model architecture, training efficiency,
multilingual intelligence and multimodal systems.

Korean
English
Code
Reasoning
Instruction Following

Attention
Tokenization
Efficient Training
Memory Efficiency
Inference Optimization

Vision Understanding
Speech Recognition
Speech Generation
Image Generation
Cross-modal Reasoning

Model Weights
Technical Reports
Datasets
Benchmarks
Experiments

# How SOVYN Thinks About Scale

Bigger is not the only direction.

We explore whether architecture, data quality, tokenization and training
can produce stronger models without requiring massive compute budgets.

01 Architecture Better use of every parameter.

02 Data Cleaner and more useful training data.

03 Training Efficient optimization and experimentation.

04 Inference Models that can run outside large datacenters.

# Research Principles

01 BUILD Start from the architecture, not from the hype. Experiment with model design, training and data directly.

02 MEASURE Document parameters, training behavior and limitations. Research should be reproducible.

03 RELEASE A model becomes more useful when people can inspect, run and build on it.

# SOVYN Ecosystem

Models Pretrained and instruction-tuned SOVYN models.

Datasets Training and evaluation datasets.

Research Architecture experiments and technical reports.

Spaces Interactive demos and model experiments.

   

# Development Status

SOVYN 1 85M

SOVYN 2 300M

SOVYN 2.3 Fast ~1B

Current work is focused on the SOVYN 2.3 generation.

# Project Meridian

SOVYN 2.3 Fast
│
├── Language
│   ├── Korean
│   ├── English
│   └── Code
│
├── Architecture
│   ├── GQA
│   ├── QK-Norm
│   ├── RoPE
│   ├── SwiGLU
│   └── Local / Global Attention
│
├── Efficiency
│   ├── BF16
│   ├── Gradient Checkpointing
│   ├── 8-bit Optimizer
│   └── Tied Embeddings
│
└── Omni Research
    ├── Vision
    ├── Speech
    ├── Image Generation
    └── Speech Generation

# Open Research

SOVYN aims to publish model artifacts whenever they are ready
and when licensing and technical conditions allow.

Weights Model checkpoints and releases.

Configs Architecture specifications.

Data Dataset sources and documentation.

Eval Benchmark results.

Reports Technical notes.

Models built for real work.

Open models  ·  Independent research  ·  Efficient intelligence

_{SOVYN Research}