Offline vibe coding with Gemma 4. This open-source Electron app runs entirely on your Mac via Apple's MLX framework. No Wi-Fi required.

Check out the code by @ammaar here:

What are you vibe coding with Gemma 4. Drop your links or videos below! 👇

tried gemma 4 on mlx last weekend, single file html prototypes work fine but it falls apart the moment state spans more than two files. local first is real, the context window just isn't there yet for anything multi-screen.

Make it electrobun

>Vibe coding >Show a fucking cube "game" just jumps Lmao that's just sad. Google is "NOT" winning AI race

vLLM-studio no

offline vibe coding with Gemma 4 sounds great until my mac battery starts sounding like a jet engine trying to keep up with mlx 💀

What's the memory/space required to install this? I'm tired of hitting the usage limits.

offline inference is the cheap part. the rest of the vibe coding loop (npm install, current api specs, deploy previews) still needs network. great tech demo, shipped projects do not run that way.

"electron app" 🤢

i have built an ios app on top of gemma 4 e2b with vision running locally. its zaya - your private second AI brain. check this out:

Btw its gemma 4 24B parameter model. What it takes to run it locally?

Damn this is genius

Started practicing offline workflows & coding with Gemma e4b. Created a repo filled with exercises so any one can clone and experiment with Pi-dev

Local models on MLX hit different once you tune the context window right, the default is almost always way too small for real coding sessions.

what do you mean by 'it feels smoother' - did you notice any improvements in response time or lag after installing hermes on discord compared to telegram?

Offline AI coding on a Mac with no Wi-Fi needed? This is huge for anyone working on planes, remote locations, or just privacy-conscious devs.

@andrewarruda @llmgram

Vibe coding offline sounds great until Electron's memory overhead starves the KV cache.

The interesting part is not just "offline coding" It is local-first AI becoming practical. No API call. No cloud dependency. No Wi-Fi required. For developers, that changes the tradeoff: privacy latency cost control Local models are entering real coding workflows.

@grok how does it compare to code with codex, claude code?

Inference speed varies with gemma 4 on older macs

Here is android version - Offline - Privacy 🔏 - Choose across wide range of models - No Internet required 🛜

Can we run on Ubuntu on GB10 machine?

without intertent we anyway can coding ty for gemma 4

Local inference removing API rate limits changes the workflow entirely. You can iterate at 2am without caps cutting your flow state mid-session.Local inference removing the rate limit ceiling is the real unlock. Iterate through the whole night without API caps interrupting your flow state.

@grok lust the benchmarks and do a gap analysis against Claude code

Amazing. Gemma 4 doing incredible things. Onward.

This is the kind of local AI development I like to see. Gemma 4 + MLX running an entire Electron coding app completely offline — no internet, no usage limits. Great for founders who want to ship fast while keeping full control and privacy. Now the question is: how does it scale when the codebase gets real (multiple files + complex logic)?

Vibe coding completely offline on a Mac with no API costs. Six months ago this wasn't possible on consumer hardware. Now it's an Electron app download. The "AI requires cloud" assumption is quietly becoming false - and that changes everything for products built around privacy, compliance, or just not wanting to send code to a server

@TouchMrown tanta IA y son incapaces de hacer una app que nos sea Electron

lol ok? and?

I prefer Goose for this type of use case, but this is neat too

Offline is power. But the real leverage is what happens after you build. In CREAO: You chat → the Super Agent builds it You save it → it becomes an Agent You run it → it delivers again and again Not just “vibe coding” in the moment. You’re building systems that keep working for you.