## TL;DR
The open-source AI movement has democratized access to frontier models. By 2025, open-weight models (Llama, DeepSeek, Mistral, Qwen) match or exceed GPT-4 on key benchmarks, transforming AI from a centralized service to a distributed ecosystem.

## Core Explanation
Open-weight spectrum: fully open (Apache/MIT — DeepSeek-V3, Mistral), community-licensed (Llama — free with restrictions), and open-ish (weights available, training data not). Hugging Face hosts 500,000+ models. vLLM and Ollama enable local inference on consumer hardware.

## Detailed Analysis
DeepSeek-V3 achieved GPT-4o parity at ~5% of the training cost via architectural efficiency (Multi-head Latent Attention, DeepSeekMoE). This challenged the narrative that only companies with billion-dollar compute budgets could build frontier models. Qwen (Alibaba), Mistral Large (France), and Gemma (Google) further diversify the ecosystem.

## Further Reading
- Hugging Face Open LLM Leaderboard
- Epoch AI: Compute Trends
- AI Index Report: Open vs Closed Models