Open weights have quietly become the default for privacy-sensitive work, leaving the closed frontier for the few problems that truly need it.
For reasoning: DeepSeek R1. For general chat: Qwen3-235B-A22B (big but open) or Qwen3-32B for smaller boxes. For coding: Qwen2.5-Coder-32B. For tiny devices: Qwen3-1.7B, Phi-4-mini, Gemma 3n.
For Europe and multilingual work: Mistral Large 3. For Chinese-heavy workloads: Qwen2.5-72B. For long documents: Kimi K2 and Llama 3.3 70B with YaRN.
Hardware reality check: 32B models need roughly 24GB of VRAM or a beefy CPU with tons of RAM. 70B models need multi-GPU or a Mac Studio with unified memory.
The best part: everything here works offline, costs nothing per token, and never writes your prompts to anyone's log. That's why open weights won where they matter most.