The public Artificial Analysis board is mostly hosted APIs. Local still wins on privacy, air-gapped shops, and predictable unit economics.
Why local
- Data never leaves the box
- No per-token surprise bills
- Latency you can measure
Hardware sketch
| Class | VRAM (rule of thumb) |
|---|---|
| 7–8B | ~8GB |
| 13–14B | ~16GB |
| 30–34B | ~24GB |
| 70B-class (quantized) | 24–48GB |
Stack
Ollama or LM Studio for the boring path. Pull a current open instruct/coder build (Llama / Qwen / DeepSeek coder lines move often — pin versions). Quantize sensibly; Q4/Q5 is still the practical default for single-GPU boxes.
Honest comparison
Local open weights rarely sit at #1 on the AA intelligence index next to Opus 5 / GPT-5.6 Sol. That’s fine. Use cloud frontier for hard agentic work; use local for private drafts, RAG over sensitive corpora, and offline coding.
Board for the hosted side: /leaderboard.