Meta shipped Llama 4 in January. Mistral followed with Large 3 in March. DeepSeek's V4 dropped in April. By the end of Q1 2026, three open-weights models were within five benchmark points of GPT-class frontier models on most standard tasks — coding, multilingual reasoning, structured output.
The self-hosting case suddenly makes economic sense for any team doing more than $50K/month in inference. A two-H100 deployment running quantised Llama 4 70B handles roughly 200 concurrent users at human-readable latency, for a hardware cost that pays back in three months at frontier API rates.
What self-hosting still doesn't buy you is the data privacy story you assume it does. The training data of an open-weights model is not, by default, auditable. If your compliance posture requires 'we can prove no client data was used in training,' you need a model trained from scratch on a curated corpus — and that is a different cost structure.
The one frontier capability still missing in open-weights is long-horizon agentic behaviour. Llama 4 and DeepSeek-V4 are excellent question-answer models. They are not yet excellent multi-tool-call planning models. If your product is RAG or chat, self-host. If your product is autonomous agents, the frontier APIs are still earning their keep.
We're now defaulting clients to a hybrid posture: open-weights for everything customer-facing, frontier API for internal automation. The savings are immediate; the engineering effort is two sprints at most.
