What it is
The trend of running coding models on your own hardware (laptop, workstation, server) instead of a cloud API. Open models (Qwen Coder, DeepSeek, Devstral, GLM, Kimi) have become strong enough, and tools like Ollama simple enough, that a local coding agent is now realistic.
Where it came from
- 2024–25 — open code models (Qwen 2.5 Coder, DeepSeek-Coder) caught up with mid-tier commercial ones.
- Ollama, LM Studio and llama.cpp turned running a model into a one-liner.
- Combinations of OpenCode/Cline/Continue plus a local model produced a full agent with no cloud.
Why it took off
- Privacy: code and data never leave the machine — critical for regulated industries (44% of companies name data privacy as their main concern when adopting LLMs).
- Economics: no token bills, no rate limits, works offline.
- Affordable hardware: about 24GB of VRAM is enough for Qwen 2.5 Coder 32B (Q4) at a quality close to top cloud models for many tasks.
How to use it today
- Install Ollama, pull a current Coder model, hook it up to Cline/Continue/OpenCode.
- Keep sensitive code and private repos local; send only what you don't mind exposing to the cloud.
- Hybrid: light tasks locally, heavy reasoning in a top cloud model.
What to watch out for
- Local models still lag the best cloud ones on hard agentic tasks — check the benchmarks.
- You will hit a hardware wall: big models need VRAM; on a weak machine you trade quality or speed.
- "Local equals safe" only holds if you control where the weights came from (weight supply chain).