Why and How to Run Local Models in Zed (zed.dev)

Zed has announced a significant rise in the usage of local AI models, reporting a threefold increase within the last 10 weeks. This surge highlights a growing preference among developers for local models, which offer enhanced privacy, reduced costs, and increased control over AI workflows. By utilizing open-weight models and innovative tools like LM Studio, Ollama, and llama.cpp, users can run capable models on their local machines without relying on cloud services that may alter pricing or limits unpredictably. Local models, such as the Qwen 3.6 35B A3B—a mixture of experts model—allow developers to maintain data sovereignty while reducing costs related to cloud computing. The innovative MoE architecture means that while the model contains 35 billion parameters, only a small subset is activated during processing, improving computational efficiency. Despite the advantages, local models do impose limitations such as reduced intelligence and context window size compared to frontier models. Developers are encouraged to experiment with various configurations to optimize performance and maximize the capabilities of local models, fostering a more resilient and adaptive AI environment.