Choosing a Local Model

Running local models with kew's Lemonade runner offers several advantages: - $0 per run — no API costs - Nothing leaves the machine — complete data privacy - Wall-clock and watts, not dollars — the honest cost axis

Sizing to Your RAM

• ~7–9B models fit comfortably on 16GB RAM
• 30B MoE (e.g. Qwen3-Coder-30B-A3B, ~3B active) fits ~32GB and is faster than a dense 30B
• Dense 30B+ models require more RAM

Pick a Coding-Tuned Model

Tool-calling reliability matters most. Coder-tuned models (Qwen2.5-Coder, Qwen3-Coder) emit well-formed tool calls far more reliably than general models.

Set an Adequate Context Window

The default context window of 4096 tokens is often too small for kew's prompts and file analysis. See Local Runner Troubleshooting for how to increase it.

The Honest Cost Axis

Local inference costs are measured in wall-clock time and watts, not dollars. This makes it easier to reason about performance trade-offs.