local vs api
You’re choosing whether the agent runs on your machine or through an API—privacy and offline access versus long-context horsepower.
recs
| model | best when | avoid when | cost vibe |
|---|---|---|---|
| Local mid/strong (qwen/llama/deepseek-class) | The code is private, the work must stay offline or air-gapped, and predictable unit cost matters. | The job needs huge multi-file context or your local tool harness is weak. | ~free–$$ (gpu) |
| API Sonnet-class | The repo is messy, the context is long, or the agent needs a tool-heavy loop. | The work has pure privacy requirements. | $$$ |
| API fast/cheap | You’re running high-volume, simple jobs from the cloud. | The mandate is “just make it local.” | $ |
why
- Local wins on privacy, latency to disk, and cost at volume—but loses when context and tool discipline matter.
- API frontier models still win on messy agent loops and fat repos.
- Hybrid is real: use local for scaffolding and tests, then an API for the hard pull request.
- Don’t pick “local” as an identity. Pick it as a constraint.