Models

This section collects model-level deep dives: how a model is built, what its architecture terms mean, how it was trained, and what it takes to run it (API or self-hosted).

Executive summary

Use Models when you need more than a benchmark table — architecture, training pipeline, licence posture, and practical deployment shapes for open-weight and closed-source systems.

When to use this page

• Comparing open-weight vs closed API models for a client or product decision.
• Preparing a technical briefing on a specific model (architecture, MoE, context, quantisation).
• Planning self-hosted inference or understanding why a model needs data-centre hardware.
• Pair with the Learning Map, Model FinOps and Articles.

Start here

Guide	Focus
AI model landscape 2026	Open source vs open weight vs closed; leading models by category; selection and enterprise evaluation
🤗 Transformers	Credited learning path through Hugging Face Transformers — install, infer, train, quantise, tasks and flagship models
Kimi K3	Deep dive — architecture, training, terminology and local deployment

Categories

Category	What you will find	Start
Open source	Open-weight / downloadable checkpoints, licences, self-hosting and architecture guides	Open source
Closed source	Hosted API models, vendor contracts, and deployment via provider platforms	Closed source
🤗 Transformers	Playbook articles adapted from the official Hugging Face Transformers docs (with credit)	Transformers

Currently covered

Model	Category	Focus
GPT-5.6, Claude, Gemini, Grok, Nova	Closed source	API contracts, routing, tools, safety, production engineering
Kimi K3, Qwen 3.5, Gemma 4, Llama 4, DeepSeek, Mistral, GLM-5, Nemotron 3, Command A+, gpt-oss	Open source	Architecture, licences, training and self-hosted deployment
BERT, GPT-2, T5, Llama, Mistral, Whisper, ViT, CLIP, Qwen2, Phi	🤗 Transformers flagships	Architecture summary, usage and links to official HF model docs

Engineer briefings for every family: LLM technical reading map.

Continue with the landscape guide, 🤗 Transformers or Open source overview.

Key takeaways

• Open-weight ≠ free of licence constraints; always read the model licence.
• Open-weight ≠ easy to run locally at frontier scale — storage, HBM and interconnect dominate.
• Closed-source models trade control for operational simplicity and vendor SLAs.
• Hugging Face Transformers is the shared model-definition layer for most open checkpoints — see the credited Transformers section.

What to do next

Read the AI model landscape 2026, browse Open source and Closed source, or follow the LLM technical reading map.