Model Training in Progress

Language Models
Beyond Attention

We're building a fundamentally new architecture for language understanding. No attention mechanism. No tokenizer. Pure linear complexity — processing raw bytes at unprecedented scale.

O(L)
Complexity
500M+
Parameters
256
Vocabulary
Zero
Attention Layers
Not an optimization.
A departure.

Others are making attention faster. We removed it entirely — and built something new from first principles.

Conventional

Transformer-Based Models

  • Quadratic O(n²) complexity — cost explodes with context
  • Tokenizer-dependent — language bias baked in
  • Fixed architecture — manually designed, static capacity
  • Context window capped by memory constraints
Our Approach

Attention-Free Architecture

  • Linear O(L) complexity — scales without cost explosion
  • Byte-native — processes raw UTF-8, zero language bias
  • Self-growing — model expands its own architecture during training
  • Theoretically unlimited context from state-based design
// Transformer
attention: O(n²) quadratic
tokenizer: required (30K-100K vocab)
context: bounded by memory

// Dreamera
attention: none
tokenizer: none (raw bytes, vocab=256)
context: theoretically unbounded
complexity: O(L) linear
Core Capabilities

Each component is designed to solve a fundamental limitation of current language models.

Attention-Free Computation

Not sparse attention. Not linear attention. No attention mechanism at all. A completely different approach to sequence modeling that processes information in constant time per step.

Byte-Native Processing

Operates directly on raw UTF-8 bytes with a vocabulary of just 256. No tokenizer means no language bias, no information loss, and no vocabulary mismatch across languages.

Linear Scaling

Time and memory complexity scale linearly with sequence length. Double the input, double the cost — not quadruple it. This makes long-context processing economically viable.

Self-Growing Architecture

The model autonomously expands its own capacity during training — adding depth and parameters exactly when needed, without manual architecture search.

Universal Multilingual

Byte-level processing treats every language and script identically from the ground up. Korean, English, Arabic, CJK — all first-class citizens by design.

Unbounded Context

State-based architecture enables theoretically unlimited context. No sliding windows, no chunking, no retrieval augmentation needed.

Building the future of language AI.

We're a Seoul-based research lab. Our model is actively training.
Interested in what we're building? Let's talk.

dreamera.co.kr@gmail.com →