Findings for running a local coding agent (Pi CLI, via Ollama's OpenAI-compatible endpoint) at six memory budgets: 8, 12, 16, 20, 24, 28 GB.
- Total memory (model weights + KV cache at the listed context) fits the budget
- Mixture-of-experts (MoE) preferred over dense
- Multi-token prediction (MTP) preferred, if available
- Must run in Ollama