Codex + Gemma 4: 100% Free, Local & Unlimited Coding Via Ollama

Añadido: 2026-05-19

[00:00:00]Codex has quietly become one of the most talked about coding tools in the last few months.

[00:00:05]OpenAI shipped it as a desktop app for Mac, then Windows support followed, and the team has been pushing updates almost weekly since. We're talking new features like the in-app browser for visual editing, parallel agent runs, code review through GitHub tags, the agents.md system for project context, and a steady stream of model upgrades.

[00:00:27]3 million people are using Codex every single week as of April, which is a 5x jump in just 3 months. Sam Altman even posted about resetting Codex usage limits every time the product crosses another million users until it hits 10 million weekly. The reason that growth matters is because Codex is now in a real fight with Claude Code. For a while, Anthropic owned the agentic coding space. Claude Code was the default answer when someone asked which terminal AI coding tool to use, but Codex caught up fast. The new GPT-5.5 model is performing strong on coding benchmarks, and the desktop app gives it a polish that Claude Code doesn't really match. You get work tree visualization, terminal tabs, preview panes, MCP server support, and a browser that can load local servers, and let you annotate elements directly. The agentic loop feels tight. Tasks that used to need five manual steps now run in one prompt.

[00:01:25]And the pricing comparison with Anthropic was so direct that OpenAI literally launched a $100 pro tier in April to match Claude Max dollar for dollar. Speaking of pricing, this is where things get interesting. Codex sits behind a few paid walls right now.

[00:01:42]The cheapest entry is ChatGPT Plus at $20 a month, which includes a small Codex allowance that most serious users blow through within a week. Then you have the new pro tier at a $100 a a which gives you five times more Codex usage than plus.

[00:01:59]The big one is the $200 pro plan with 20 times the plus limit, designed for full-time engineers running parallel projects. OpenAI also moved to token-based billing this year, replacing the older per message model. So, your actual cost now depends on how heavy your prompts get, how long your sessions run, and which model you pick. Solo devs and hobbyists feel this pretty quickly.

[00:02:23]The way around it is Ollama. If you haven't used Ollama before, it's basically a local runtime for open-source models. You install it, pull a model from the registry, and it exposes an OpenAI compatible API on your machine at localhost port 1434.

[00:02:40]That second part is the whole trick.

[00:02:42]Because Ollama mimics OpenAI's API format, Codex can talk to it like it's talking to OpenAI's servers, except nothing leaves your hardware. And the Ollama library has hundreds of models.

[00:02:54]Gemma 4, DeepSeek, Quin 3 Coder, Llama 4, Mistral, GPT-OSS from OpenAI itself, Kimmy K2, GLM 4.7. Whatever you want, you can swap models in and out with a single command, and Codex will use whichever one you point it at. So, the setup I'm showing today works the same way for any of those models.

[00:03:16]For this video, I'm going with Gemma 4, and it's the strongest open release Google has ever shipped. The biggest shift is the license. Earlier Gemma versions used a custom Google license with restrictions and monthly active user caps, which made enterprise legal teams hesitant. Gemma 4 ships under Apache 2.0. Full commercial freedom, no caps, no acceptable use enforcement. You can build whatever you want on top of it and ship it. The family comes in four sizes. The E2B is the smallest, around 2 billion effective parameters, runs in under 1.5 GB of memory, designed for phones and lightweight hardware.

[00:03:57]Then E4B at about 4 billion effective parameters for edge devices and laptops.

[00:04:03]After that, you have the 26B mixture of experts model that only activates 3.8 billion parameters at a time, which is sweet for consumer GPUs. And finally, the 31B dense model, which currently sits at number three on the Arena AI leaderboard among all open models with the 26B MoE right behind it at number six.

[00:04:25]The E in E2B and E4B stands for effective parameters, by the way. Google uses a technique called per-layer embeddings, which inflates the parameter count on paper, but uses way less compute during actual inference. So, the small ones punch above their weight class significantly. They run offline, handle text, images, and up to 30 seconds of audio input through a conformer encoder. All four Gemma 4 models are natively multimodal from the architecture level, not as a fine-tuned add-on. Context window goes up to 256K tokens. Multilingual support across 140 plus languages. Native function calling and structured JSON output, which is exactly the kind of behavior you want for agentic coding. The 31B scores 89.2% on AIM 2026, and its Codeforces Elo sits around 2050, which is expert competitive programmer level for something you can run on a single GPU.

[00:05:24]Okay, on to the actual setup. First thing you need is Codex itself, or you can use the Codex CLI through NPM if you prefer terminal only. Either one works with Ollama. Next, install Ollama. Grab the installer from their site, or use the curl command in your terminal.

[00:05:43]Make sure you're on version 0.24 or newer because the Codex launcher integration needs the latest build. Run Ollama in the background. You should see the small icon in your menu bar or system tray confirming it's live. Now to pull the model, open your terminal and type Ollama pull Gemma 4 colon E2B. Hit enter and you'll see the download start.

[00:06:05]For E2B, it's a small file, only a couple gigabytes. If you have more VRAM and want better output, swap E2B for E4B, which is closer to 9 gigabytes.

[00:06:16]The 26B and 31B variants need a real GPU with 16 to 24 gigabytes of VRAM, depending on quantization. If you're unsure what your hardware can run, there's a site called can I run AI locally, where you plug in your specs and it tells you exactly which models are realistic. I'm staying with E2B today because I want to prove that even the tiniest variant produces usable output through Codex.

[00:06:42]Once the model finishes downloading, you can test it directly. Run Ollama run Gemma 4 colon E2B and type a quick prompt to make sure it responds. If you see output, you're set. Now to wire it into Codex. The clean way is the one-line launcher Ollama added. In your terminal, run Ollama launch Codex. This command opens a picker that detects every model you have installed locally and shows them in a list. It also shows Ollama's cloud recommended models like GLM 4.7 flash, Kimiko 2, and Quen 3 coder. But those need an Ollama cloud subscription, so I'm skipping them.

[00:07:20]Select Gemma 4 E2B from the list, hit enter, and the Codex app launches with that model wired in as the active provider.

[00:07:28]Once Codex opens, you'll see a small Ollama tag indicating local mode is active. From here, the whole Codex experience works the same as it would with GPT-5 Codex. You can prompt it for agentic coding tasks, let it edit files in your working directory, use the built-in browser to load local sites and annotate elements, run terminal commands, all of it. The only difference is the model intelligence underneath and the response speed depending on your hardware.

[00:07:55]For the demo, I'm going to give it a front-end task. I'll ask it to build a clean SAS landing page in a single HTML file. Hero section, feature grid, pricing block, call to action. Just a plain prompt, no extra instructions. Let me hit send and see what comes back.

[00:08:12]Okay, it's working through the request.

[00:08:14]On my laptop, it's showing around 30 to 40 tokens per second, which is plenty fast for iterative coding.

[00:08:21]Not API speed, obviously, but you don't sit around waiting, either.

[00:08:25]Output is done. Let's see the preview.

[00:08:28]And there we go. The landing page looks clean. Gradient hero, three-column feature section, simple pricing cards, footer with social links. CSS is inline.

[00:08:38]The markup is semantic. Responsive breakpoints are present. For a 2 billion effective parameter model running offline on consumer hardware, that's a real result. Not Claude Opus quality, sure, but for a free local setup with no token meter, this is a workflow you can actually use day-to-day. One last thing.

[00:08:58]If you ever want to undo the Ollama integration and go back to regular Cloud CodeX, run Ollama launch codex {dash} {dash} restore in your terminal. That resets the config and you're back on your chat GPT plan. The bigger takeaway is this. Open models like Gemma 4 plus a runtime like Ollama plus the CodeX desktop app gives you a fully agentic coding stack that runs on your own machine, private, free, no rate limits.

[00:09:26]And the model swap is one command. So, when something better drops next month, you pull it and keep working. Links to the Gemma 4 model card, Ollama CodeX docs, and OpenAI CodeX pricing are in the description if you want to dig deeper. All right, so that's it from the video and I hope you enjoyed it. If you did, please like this video and subscribe to the channel and I'll see you in the next video.