Self-hosted inference is back. We mapped which SaaS products are running their own Llama, Mistral, or fine-tuned models in production.