Llama 4 from Meta is the open-weights flagship. Quality on the 405B variant approaches GPT-4-class on most benchmarks, and the open license means you can self-host with no per-token cost.
Hosted access via Together / Groq / Fireworks comes in cheap at ~$0.0008/1K tokens equivalent. Tool use is improving but lags closed models.
For regulated industries, privacy-sensitive workloads, or sheer cost optimization at scale, Llama 4 is the obvious pick.
Verdict: best open-weights LLM. 8.2/10.