I replaced my Gemini subscription with a local 4B model for research, and it's shockingly capable

For the average user, most of the usage of a Google AI Plus subscription is research-related. Most of the queries made to Gemini were repeated lookups, with a few common requests. Local models have a reputation for struggling with recent information, as they only know what they picked up during training. However, this changed when the author switched to a local 4B model, Gemma 4 E4B.

Gemma 4, built on the same research as Gemini 3, runs on phones and laptops, fitting comfortably on a 4GB VRAM GPU. It uses about 4.5B effective parameters and offers native tool calling, which allows it to decide when to use outside tools like the web or filesystem. This tool calling is its strongest skill, with Google's own benchmarks showing E4B scoring 42.2% on Tau2, a test for multi-step tasks requiring tool use, compared to 16.2% for Gemma 3 27B.

Gemma 4's knowledge cutoff is January 2025, but it was only released in April 2026, meaning information after early 2025 isn't available. Despite this, the model handles current information well when the web is turned on, admitting when it doesn't have the up-to-date information. Web search, a simple toggle in LM Studio and Noema, provides clickable citations at the end of web-based responses. While not perfect, this feature offers better answers when asking one clear question at a time.