Open vs. Closed AI Models

What does "open" mean for AI models?

The AI industry uses "open" differently than traditional software. A spectrum exists:

• Closed/Proprietary: GPT-5.1, Claude 4.5. You can only access them through APIs. The weights, training data, and architecture details are secret.
• Open weights: Llama 3.3, Mistral. You can download the trained model and run it yourself. But the training code and data are often not released.
• Fully open: Some research models release weights, training code, and data. Rare at the frontier.

Most "open-source AI" is actually open-weights: you get the result of training, not the recipe to reproduce it.

What can you do with open-weights models?

With Llama or Mistral, you can:

• Run locally: No API calls, no usage fees, complete privacy
• Fine-tune: Adapt the model for your specific domain or task
• Modify: Remove safety filters, change behavior (with all the responsibility that implies)
• Deploy anywhere: Your servers, your cloud, your rules
• Inspect: Study how the model works, run interpretability research

What are the trade-offs?

Open models offer:

• Control over your data (nothing sent to external servers)
• Customization (fine-tune for your use case)
• Cost predictability (hardware costs, not per-token fees)
• Independence (no API changes, no service shutdowns)

Closed models offer:

• State-of-the-art capability (GPT-5.1, Claude 4.5 still lead)
• No hardware management
• Continuous improvements (providers update models behind the API)
• Safety infrastructure (moderation, filtering, monitoring)

The major open model families

• Llama (Meta): The flagship open model. Llama 3.3 405B approaches frontier closed model capability. Permissive license for most uses.
• Mistral (Mistral AI): French company, strong models, competitive with larger Llama variants. Some models fully open, some commercial.
• Qwen (Alibaba): Strong multilingual performance, especially Chinese. Various sizes and specializations.
• Gemma (Google): Smaller models for research and development. More restricted license than Llama.
• Phi (Microsoft): Small but capable, designed to prove that smaller models can perform well.

Running open models yourself

The ecosystem for running open models has matured:

• Ollama: One-command setup for running models locally on Mac, Windows, Linux
• llama.cpp: Efficient C++ implementation that runs on consumer hardware
• vLLM: High-performance inference for server deployments
• Text Generation WebUI: Browser interface for local models

With a decent GPU (16GB+ VRAM), you can run 7-13B parameter models comfortably. For larger models, you need multiple GPUs or cloud instances.