Models vs Providers

Models determine capability: how smart the AI is, how fast it responds, and what kinds of work it can do.

Providers determine operating constraints: privacy policy, legal jurisdiction, availability, and access model.

Models

Models are the AI systems themselves. In practice, it helps more to think in capability classes than in winner lists.

Model class	What it is best at
frontier reasoning models	complex agentic work, debugging, and multi-file changes
strong general-purpose coding models	everyday implementation and review loops
multimodal models	screenshots, mockups, diagrams, and visual/UI work
long-context models	large investigations when context is still selective
hosted open-weight models	more control without going fully local

Visual/UI Tasks

Use a strong multimodal model for screenshots, mockups, and design files. That is usually the right fit for implementing UI from designs, debugging visual issues, or working from component libraries.

Providers

The same model can be available through multiple providers:

Model	Direct Provider	Also Available Via
Claude Sonnet 4.6	Anthropic	AWS Bedrock, Google Vertex
GPT-5.2	OpenAI	Azure OpenAI
Llama 4	Meta (weights only)	AWS Bedrock, Azure, Together AI
Mistral Large	Mistral	AWS Bedrock, Azure, Google Vertex

Why This Matters

For Privacy

Different providers have different data policies:

• Anthropic direct: US jurisdiction, 30-day retention
• AWS Bedrock: Your AWS account, your retention settings

For Enterprise

• Azure OpenAI: EU data residency available
• AWS Bedrock: Private endpoints, VPC integration

For Workflow

• Direct providers: best when you want the vendor’s native tooling and latest releases
• Aggregators: useful when you want one key, fallback routing, or easy model switching
• Cloud platforms: useful when procurement, networking, or compliance requirements matter more than raw convenience

Aggregators & Cloud Platforms

Platform	Type	Models	Why Use It
OpenRouter	Aggregator	200+	One key for all models, fallback routing
Together AI	Aggregator	100+	Open models, fine-tuning
AWS Bedrock	Cloud	Claude, Llama, Mistral	Enterprise, VPC integration
Azure OpenAI	Cloud	GPT, DALL-E	EU regions, compliance
Google Vertex	Cloud	Gemini, Claude	GCP integration

Open Weight Models

“Open weight” means the model weights are publicly available, so you can download and run them yourself. That is not the same thing as open source, which would also include training code and data.

Model	Params	Why It Matters
Llama 4	405B	Meta’s flagship, runs locally or hosted
DeepSeek V3.2	671B MoE	Strong open-weight coding model
Qwen 3	235B	Strong multilingual, tool use
Kimi K2.5	1T MoE	Moonshot’s flagship, massive scale
GLM 4.7	320B	Zhipu’s latest, strong reasoning
Mistral Large	123B	EU-based, good reasoning

Why Care?

Self-hosting: Run on your own hardware. Zero data leaves your network and provider boundaries stay under your control.

No vendor lock-in: Switch providers freely. Model weights are yours.

Fine-tuning: Train on your codebase for domain-specific improvements.

Offline: Works without internet. Air-gapped environments.

The Tradeoff

Open weight models lag ~6-12 months behind frontier closed models (Claude, GPT). For most coding tasks, this gap is negligible. For cutting-edge agentic workflows, closed models still lead.

Running Locally

Tool type	What it does
local runner	one-command local inference for individuals
desktop runtime	GUI-based local model management
serving stack	production-grade local or self-hosted inference

Pair with the editor or terminal workflow you already use. Note: “free” local models still require capable hardware (16GB+ RAM, GPU with 8GB+ VRAM recommended).

Next Steps

• Capabilities Matrix: compare what models can do
• Selection Guide: choose by workflow, privacy, and capability