AI is changing the world. Don't stay behind. Clear summaries, community insight, delivered without the noise. Subscribe to never miss a beat.

Privacy

Contact

Back to all news

llms google-gemma claude developer-tools

Running Gemma 4 locally with LM Studio's new headless CLI and Claude Code

Running Google Gemma 4 Locally With LM Studio’s New Headless CLI & Claude Code

ai.georgeliu.com

April 5, 2026

20 min read

🔥🔥🔥🔥🔥

55/100

Summary

LM Studio 0.4.0 introduces llmster and the lms CLI for running Google Gemma 4 26B locally on macOS. Local inference provides advantages such as avoiding cloud API rate limits, reducing costs, enhancing privacy, and minimizing network latency.

Key Takeaways

LM Studio 0.4.0 introduced a headless CLI and the llmster inference engine, allowing local model execution without a graphical interface.
Google’s Gemma 4 26B model utilizes a mixture-of-experts architecture, activating only a fraction of its parameters per forward pass, enabling efficient local inference on standard hardware.
The Gemma 4 model family includes variants optimized for different hardware, with the 31B dense model achieving high benchmark scores while the 26B-A4B model offers competitive performance with lower memory requirements.
Running local models eliminates issues related to cloud AI APIs, such as rate limits and privacy concerns, providing cost-effective and consistent availability for tasks like code review and prompt testing.

Read original article

Community Sentiment

Mixed

Positives

The introduction of a headless CLI for running Gemma 4 locally enhances accessibility for developers, enabling more users to leverage advanced AI capabilities.
Using Claude Code as a frontend for Gemma 4 indicates a growing interest in user-friendly interfaces for AI models, which could lead to broader adoption.

Concerns

Anthropic's cautious approach to updates suggests a reluctance to fully embrace broader applications of their models, potentially limiting innovation and flexibility for users.
The clarification on MoE's memory usage highlights a misconception about its efficiency, indicating that users may not fully understand the implications of model architecture on resource consumption.