architecture-catalog.txt
artificial-intelligence/large-language-models/models/architecture-catalog.txt · 48 line(s) · txt
SLeeLa LLM MODEL TAXONOMY โ QUICK REFERENCE
CLASSICAL / STATISTICAL
- N-gram language models
- Hidden Markov models
- Conditional random fields
NEURAL SEQUENCE MODELS
- Feed-forward neural language models
- RNN, LSTM, GRU
- Sequence-to-sequence and attention-based encoder-decoder
TRANSFORMER FAMILIES
- Encoder-only (BERT-style)
- Decoder-only autoregressive (GPT-style)
- Encoder-decoder (T5-style)
- Sparse mixture-of-experts
- Long-context variants
ALTERNATIVE / HYBRID SEQUENCE MODELS
- State-space language models
- Hybrid attention/state-space models
- Neuro-symbolic systems
MULTIMODAL / SPECIALIZED
- Vision-language
- Speech/audio language models
- Code language models
- Multimodal text-image-audio-video systems
SYSTEM AND TRAINING PATTERNS (NOT ARCHITECTURES BY THEMSELVES)
- Retrieval-augmented generation
- Tool-using / agentic systems
- Continual learning
- Federated training
- Quantization
- Knowledge distillation
RESEARCH DIRECTIONS
- Diffusion-language hybrids
- Neuromorphic language systems
TIME HORIZON
The catalogue represents families documented through 2025. The years 2026โ2500 are a speculative horizon, not a factual inventory of known models.
CAUTION
These categories overlap. A deployed model can belong to several categories simultaneously. This is a representative taxonomy, not 2,000โ2,500 separate model entries.