Transcription
Gemini CLI just shipped its biggest update yet. Version 0.40 brings local Gemma support straight to your terminal. The Google Gemma team confirmed it on launch day.
Over 150 improvements landed in this single release. Security, memory, and a streamlined UI got serious upgrades. And the roadmap says full local model execution is coming next.
Local Gemma is experimental, but already works. You set it up with the Gemini Gemma setup command. The CLI downloads the light RT runtime and a 1 GB Gemma 3 model. You accept the Gemma terms of use, and the install runs locally. Then you check status with the {slash} Gemma command.
To turn the router on, open {slash} settings and flip the enable Gemma model router toggle. Or set experimental.gemma to true in your settings.json file.
So, what is Gemma actually doing here? Gemma 3 is a 1 billion parameter model with a GPU accelerated build. It runs on your hardware through the light RT LLM runtime. Today, it powers intelligent model routing. It decides which model handles each prompt locally. Full local model execution is next on the roadmap. Gemma 3 is the foothold.
Beyond Gemma, five upgrades stand out. Tiered memory now stores context across four distinct layers. Project rules live in a gemini.md file at your repo root. Subdirectory rules scope down to specific folders. Private memory stays local and never commits to the repo. Global memory follows you across every workspace you open. Auto memory extracts reusable skills from your old sessions automatically. The background agent drafts new skill.md files from idle conversations. Run {slash} memory inbox to review and promote what it finds.
Compact tool output replaces noisy tool call boxes with a clean structured format. File reads, searches, and directory lists all render cleanly now. Topic narrations replace agent chatter with descriptive section headers across every agent. And a task tracker keeps a persistent objective list while the agent works through complex goals.
Performance and security got serious upgrades, too. Ripgrep is now bundled into the CLI for lightning-fast offline search. Memory usage dropped significantly for dev tools and large output streams. Sandboxing improved for Docker and macOS seatbelt environments. Shell command validation blocks injection attacks before they execute. The {slash} new command starts a fresh session instantly. New color-blind friendly themes ship out of the box.
This is the local agent shift. Update Gemini CLI today and try Gemma yourself. Tried Gemini CLI already, or sticking with Claude Code or Codex?