Running Google Gemma 4 Locally with LM Studio Headless CLI and Claude Code

Gemma 4 adopts a Mixture-of-Experts (MoE) architecture, activating only 4B parameters out of a 26B model, enabling high-performance inference even on low-spec hardware. With a 256K long context, vision input, and function calling capabilities, it is optimized for local environments.

LM Studio 0.4.0 introduces a new Headless CLI called llmster. This allows model downloading, loading, chatting, and API server execution via the lms command without the desktop app. In particular, it provides an Anthropic-compatible API, enabling the setup of an environment where Claude Code can be used as a fully offline code assistant.

Users can optimize performance through the following detailed hardware tuning:

• Memory management: Predict and configure memory requirements based on context length using the --estimate-only option
• GPU offloading: Optimized utilization of Apple Silicon's unified memory and NVIDIA VRAM
• Parallel request processing: Simultaneous processing of multiple requests through continuous batching
• Efficiency improvements: KV cache memory savings through Flash Attention and automatic model unloading through TTL settings