Local models
Vane + Ollama: three things to check
Local models remove API fees, not necessarily total cost. Count hardware, speed, and maintenance.
1. What changes
You stop paying per token, but you need a machine that can run the model. Electricity, hardware depreciation, and your time become the cost. Vane connects to local models through Ollama / LM Studio and similar tools; check the official README for the exact support.
2. Rough hardware ranges
| Model size | Rough RAM/VRAM | Good for |
|---|---|---|
| 7B–8B (quantized) | 8–16 GB | Light Q&A, experiments |
| 13B–14B (quantized) | 16–32 GB | Balance of quality and speed |
| 30B+ (quantized) | 32 GB+, ideally a GPU | Higher quality, higher cost |
These are ranges, not guarantees. Quantization, context length, concurrency, CPU/GPU all change the real requirement.
3. When it is not worth it
If you only run a few dozen queries a day, have no spare hardware, and do not want maintenance, a cloud API is often simpler. Put “hardware cost ÷ depreciation months” into the calculator and compare it with API mode.
Official entry point: Vane GitHub README. This page does not provide a one-click installer.