Models, Reasoning & Inference
Explain how modern AI models operate and what system architects need to understand about them.
Why This Matters
Every production AI system sits on model behaviour: tokens, context windows, inference economics, and reasoning limits. Architects who skip this layer design agents that cannot be costed, evaluated, or bounded.
Architecture Position
Models sit at the Model Layer of the architecture map. They reason; they do not own execution, identity, or policy. See it on the architecture map.
Start Here
How LLMs Work — Complete Guide
End-to-end interactive visual overview of the entire LLM pipeline on one page.
Tokenization
BPE, SentencePiece, vocabulary tradeoffs, and how text becomes numbers.
Transformers
The transformer architecture, positional encodings, and scaling laws that power modern AI.
Inference Pipeline
Prefill/decode phases, KV-cache, speculative decoding, and serving optimization.
Core Concepts
Model Foundations
- Neural networks
- Transformers
- Attention
- Tokens
- Embeddings
Training
- Pre-training
- Post-training
- RLHF
- DPO
- Data
- Fine-tuning
Inference
- Prefill
- Decode
- KV cache
- Batching
- Quantization
- Serving
Generation
- Temperature
- Top-k
- Top-p
- Structured generation
Reasoning
- Reasoning models
- Test-time compute
- Planning vs reasoning
- Model limitations
Architecture Decisions
- Open vs closed models
- Small vs large models
- Hosted vs self-hosted
- Model selection
- Model routing
Advanced
Advanced material
Deeper guides for this domain are on the way.
Coming SoonArchitecture Patterns
Applied patterns for this domain live in the pattern library.
Browse architecture patternsCase Studies
Enterprise walkthroughs that apply this topic in a full system.
Browse case studiesRelated Topics
Context Engineering
How information is assembled, prioritized, and supplied to models and agents.
Open →Retrieval, RAG & Knowledge
How enterprise AI systems locate, rank, validate, and authorize knowledge.
Open →Evaluation, Observability & Reliability
How AI systems are measured, debugged, and operated safely in production.
Open →AI Platforms & Infrastructure
Infrastructure needed when AI moves from an app into an enterprise-wide capability.
Open →Roadmaps
This topic appears in the Enterprise AI Architect path.
Follow the Enterprise AI Architect roadmap