Unplugging completely! No WiFi and zero notifications. A great way to get deep focus on a project. Here is a walkthrough showing how to run Gemma 4 (26B A4B) fully offline with LM Studio & OpenCode to parse PDFs, ask questions, and build sites 100% locally.

Want to take your coding setup anywhere? This video, with @ianballantyne from the Gemma team, shows how to code off the grid by serving Gemma 4 on LM Studio/Ollama, connecting OpenCode via an OpenAI endpoint, using liteparse for HTML/CSS/JS, and managing context memory offline.

Had a emergency situation on the farm last week, was in an area with no cellphone signal. Urgently needed to find out what a term in documentation of animal medication ment. Realised I tested Gemma 4 on my phone, asked Gemma and got an IMMEDIATE offline answer of what the term ment, saved animal’s life 🙏

@lmstudio On your IPhone AI Agent by Gemma4 100% Open Source 100% Runs locally with no internet required, protecting privacy

The fact of the matter is that no 35 billion, 27 billion, or 26 billion model—whether quantized to Q4 or Q3 for that matter—is currently deployable on MacBooks, even if these MacBooks have compatible memory as some people are claiming. I have a MacBook M3 Pro with 36 GB RAM and I have not been able to run these models, even the Q4 GGUF models or MLX models for that matter (be it Gemma 4, Qwen3.6, or whatever). You need to have at least 64 GB RAM to get a decent output from these models. That is the bottom line.

Nice case! would like to discover local agents with Gemma4 hosted by @atomic_chat_hq they have TurboQuant implemented and MLX as well. Perfect for large context window tasks.

counterpoint: gemma 4 running locally is cool but i know for a fact i would still find a way to get distracted by a pre-installed calculator app or something 💀

The fact that a 26B model runs fully locally now tells us everything about how fast this space is moving.

What's the system requirement for this?

Or, you can just use Friday.

Nice you advertise opencode here but ban people if they use it with gemini afaik

Which one do you use in-house when using gemma as an agent? Opencode, Pi or Openhands?

From my experience Gemma4 is a maybe OK as a chat interface, but a mess as coding agent and for tool calling. Messes up tags all the time an gets hung up in infinite loops frequently. I had to revert to qwen 3.5.

I’ve noticed going fully offline really sharpens focus, this kind of setup makes that actually practical

All you need is a laptop, a solar panel, and a USB-C cable. 🗻 #MountainCoder #MountainDev #MaybeStarlinkToo

Enjoying it on iOS so far.

Local 26B still chokes on RAM. Quantize to 4-bit or disk swapping will wreck your throughput.

unplugged local runs hit different. ollama vibes all day

if i got 16gb ram, which model is perfect for me ?

🤯The Gemma 4 models are next level. The intelligence packed in is insane for the parameter counts. I think my favorite is Gemma 4 E4B because its small and scrappy. I can fit two of them in 8GB VRAM with 131k context windows using the versions we quantized to 4 bit at 30+TPS at: What's everyone elses favorite flavor of Gemma 4? Pls respond in comments 👇

So you've given up completely on Gemini CLI?

Gemma 4 is a game-changer! Before, I was required to use cloud providers, now I am able to directly integrate it into critical parts of my app that now runs locally, doing tasks that bigger models did before.

@lmstudio 如果你顺便提一下流畅运行需要什么样的硬件配置，我想大家的兴趣就没那么大了. 至少要25tps级别

finally a tutorial that doesn't require a cloud subscription to break your wifi router

offline inference is the only architecture where privacy is enforced by the hardware, not promised by the terms of service

Where can I get a no-Woke AI?

This is very interesting demo of local, edge deployment using the larger Gemma 4 26B A4B. Once I get better hardware & memory (currently 16GB), I would love to give this a try.

🔥👌🤩

Offline AI + zero distractions = real deep work unlocked

Local models + tools are getting surprisingly powerful.

Great open source model

Absolutely. A little caution goes a long way. Verifying someone's identity before sending money is one of the simplest and most effective ways to protect yourself from scams.