PewDiePie's AI: Odysseus Download, Ajax Release and PC Requirements

PewDiePie's AI is two things, and only one of them is out. Odysseus is his free, open-source AI workspace: chat, agents, deep research, email, notes and calendar in one dashboard, released on May 31, 2026, now past 88,000 stars on GitHub, and installable today with four commands. Ajax is his own model, a fine-tuned Qwen 3.5 9B with its refusals removed, announced on October 2, 2026, and not downloadable yet. Here is the part most of the coverage skipped: the "PewDiePie AI download" that half the internet is searching for does not contain an AI at all. Odysseus ships with no model inside it. You bring the brain. And the brain Ajax is built on, plain Qwen 3.5 9B, already runs on an ordinary 8 GB graphics card tonight, which means you can set up the whole thing now and swap Ajax in the day it lands.

Ethan saw the clip on Friday morning. A YouTuber with more subscribers than most countries have people was standing in front of a rack of graphics cards, saying he had built his own AI, that OpenAI had banned him twice while he did it, and that the model would run "on your own hardware". By lunch Ethan was at Jake's computer shop with his gaming laptop under his arm. "Can you put PewDiePie's AI on this? It's got a 4060." Jake, who has spent the last few months setting up local models for customers who are tired of paying monthly for chatbots, typed "pewdiepie ai download" into the search bar and found three different websites with PewDiePie's name in the address, none of them his. That is where this page starts: what Odysseus and Ajax actually are, what "uncensored" means here, what your PC needs, how to install the workspace on Windows, Linux or a Mac, and how to have the Ajax base model answering on your laptop before the real thing ships.

If running a model on your own machine is new to you, the free guide series on running AI locally takes you from zero to a working model in an afternoon, and everything on this page sits on top of that foundation. The short version: a model is a big file of numbers, a runner like Ollama loads it onto your graphics card, and a workspace like Odysseus is the window you talk to it through. PewDiePie built the window first, and the model second. Most people searching his name want the model, and that is the piece that is not out.

What is PewDiePie's AI? Odysseus, Ajax and the "council" explained

Three different things travel under the name "PewDiePie's AI", and the search results mix them up constantly. Getting them straight saves you an evening of installing the wrong thing.

The first was a private experiment. In late 2025 Felix Kjellberg, the Swedish YouTuber known as PewDiePie, posted a video titled "STOP. Using AI Right now" showing a homemade chat interface running on his own computer. It had web search, memory, document reading and voice output, and one memorable trick: a "council" of several chatbots that each answered a question and then voted on the best reply, until, as he told it, they started colluding against him. The rig behind it was absurd by home standards: ten graphics cards with 424 GB of video memory in total, eight of them RTX 4090s modified to carry 48 GB each, plus two RTX 4000 Ada cards, on a 64-core Threadripper PRO. That experiment is why "pewdiepie ai council" is still one of the most searched phrases about him. It was never released as a product.

The second is Odysseus, the public version of that idea. It went up on GitHub on May 31, 2026, collected about 74,000 stars in its first three weeks, and has since moved from his personal account to a community organization, odysseus-dev, where it sits above 88,000 stars. It is a self-hosted AI workspace: you install it on your own machine, open it in a browser, connect it to a model, and get chat, agents, multi-step web research, a blind model-comparison tool, a document editor, an email client, notes, tasks and a calendar, all in one place. It is free and licensed under the AGPL-3.0. The council lives on in it as the Compare feature, which shows answers from several models side by side without telling you which is which.

The third is Ajax, announced on October 2, 2026. It is a model, not an app: Alibaba's open-weight Qwen 3.5 9B, fine-tuned by PewDiePie to be the always-on agent inside Odysseus, handling web searches, browsing, email and calendar work offline, with its refusal behavior stripped out. His own announcement page says it is "coming soon", and that he has "decided to release Ajax when it's ready instead of putting a timer". A countdown that pointed at the morning of October 3 in Japan time was taken down the same day, so any article quoting that date is already out of date.

The three names at a glance

Keep that last row in mind, because it is the practical heart of this page. Ajax is a fine-tune, which means it has the same size, the same shape and the same memory needs as the model it started from. Anything you learn running Qwen 3.5 9B in Odysseus this week carries straight over to Ajax.

PewDiePie AI download: what you can actually get today

"pewdiepie ai download" is the single most searched phrase about this whole story, and the honest answer has two halves.

You can download Odysseus right now. It is a GitHub repository, so "downloading" it means cloning the code and starting it, usually with Docker. There is no installer with a .exe at the end, no app store listing and no account to create with anybody. The official home is github.com/odysseus-dev/odysseus. The older address under his personal account, pewdiepie-archdaemon/odysseus, forwards to the same place. Everything else, including forks with "pewdiepie" in the name under other people's accounts, is a copy that may or may not match the real thing.

You cannot download Ajax yet. As of October 2 there is no model file, no model card, no published license for the fine-tuned weights, no benchmark results and no list of quantized versions. The only official page about it is on PewDiePie's own data site, and it says "coming soon". So any page offering "Ajax GGUF download" today is offering something that is not Ajax. At best it is a renamed copy of Qwen 3.5 9B. At worst it is something you do not want running on your machine.

What you can download today, and what gets you closest to the real experience, is the model Ajax is built on. Qwen 3.5 9B is in the Ollama library under the tag qwen3.5:9b. Paired with Odysseus, that gives you the same workspace, the same tools and the same size of brain that Ajax will have. What you will not have is PewDiePie's agent training or the removed refusals. When Ajax ships, you change one model name in a settings screen.

Ethan's face fell a little at "not yet". Then Jake pointed out that the laptop under his arm could be running the same workspace, with the same family of model, before the afternoon was over, and that the day Ajax dropped, it would be a ten-minute swap instead of a weekend project. That changed the mood.

What is Ajax? PewDiePie's model, explained

Ajax is a 9-billion-parameter model built by taking Qwen 3.5 9B and training it further for one job: being the always-on agent inside Odysseus. "Agent" here means a model that does not just answer, but acts: it can search the web, open pages, read and sort your email, and manage your calendar through the tools Odysseus gives it. If the word is new, this plain-English explanation of what a language model is covers the ground underneath.

Here is what PewDiePie has said about how he built it. He spent months collecting traces of Odysseus being used: documents, email, notes, calendars, research and files, the real record of a model working through real tasks. He aimed for tens of thousands of good samples and, in his words, got "much less than expected", so he leaned more and more on generated examples that were then filtered. The model was trained on Odysseus's tool tasks with reinforcement learning, a method where the model tries a task, gets scored on whether it worked, and is nudged toward what worked. His early report is that Ajax finished its intended tasks roughly nine times out of ten. That is his own number, from his own testing, and there is no independent benchmark yet to set beside it.

Why a 9B model, and not something huge

This is the most interesting decision in the whole project, and it says a lot about who Ajax is for. PewDiePie's own machine could run models twenty times larger. He picked 9B anyway, because 9B is the size that fits on the graphics cards ordinary people own. At 4-bit compression, which costs a little quality and saves a lot of memory, a 9B model fits on an 8 GB card. At 8-bit it wants about 12 GB. Even at full precision it fits on a 24 GB card. A model designed to sit in the background all day, checking your email, has to be small enough to share the machine with everything else you do.

The base model helps too. Qwen 3.5 9B is natively multimodal, so it can look at images as well as read text, which matters for an agent that browses web pages. It is released under the Apache 2.0 license, which allows commercial use and modification. Whether Ajax inherits that license, or ships under something else, is one of the open questions. The announcement does not say.

What we still do not know about Ajax

1. The release date. "When it's ready." The countdown was removed on announcement day.
2. The license. Qwen 3.5 9B is Apache 2.0 and Odysseus is AGPL-3.0, but nothing is published for the Ajax weights themselves.
3. The formats. Coverage has mentioned GGUF for llama.cpp and Ollama, plus FP8 and full-precision versions for vLLM, but there is no official file list.
4. Benchmarks. No public test results, only the creator's "nine times out of ten" from his own runs.
5. Whether it works outside Odysseus. It was trained on Odysseus's tools. It will almost certainly chat anywhere, but its agent skills may depend on the tool formats Odysseus uses.

This list will shrink the day Ajax ships, and this page will carry a dated update when it does.

Is Ajax uncensored? What "refusals ablated" actually means

The announcement says Ajax's "refusal has been ablated, for a freer, less restricted AI experience. Please use responsibly." Headlines turned that into "uncensored AI", which is close, but the details matter if you are deciding whether to run it.

Most chat models are trained to refuse certain requests. Researchers found that this refusal behavior tends to live in a fairly narrow direction inside the model's internal math. A technique called abliteration, short for directional ablation, finds that direction and removes it, so the model stops refusing without being retrained from scratch. PewDiePie used Heretic, an open-source tool by developer Philipp Emanuel Weidmann that automates the process. It searches for settings that drive refusals down while keeping the model's answers as close as possible to the original, so it does not get noticeably dumber in the process. Anyone who can run a command-line program can use Heretic, which is a big part of why "uncensored" versions of popular open models appear within days of their release.

Two honest caveats. First, PewDiePie says he deliberately kept the lines on requests involving harm to other people or harm to the user. So Ajax is less restricted, not lawless. Second, and this is the part almost nobody is talking about: a model with fewer refusals is a different risk when it is an agent with access to your inbox. A chatbot that will write anything is a question of taste. An agent that reads your email, opens web pages and follows instructions it finds there is a question of security. Web pages and emails can hide text that tries to give the model orders, an attack called prompt injection. One of the things refusal training does, imperfectly, is make a model hesitate before doing something strange. Take that away, connect it to your accounts, and the tools you give it matter more than ever. The section on running it safely turns this into a short list of settings.

None of this is a reason to avoid Ajax. It is a reason to give an agent the same thought you would give a new employee with your passwords: what can it reach, and what happens if someone else talks it into something?

Why did OpenAI ban PewDiePie? The two bans, explained

In the announcement video PewDiePie said OpenAI banned his account twice while he was building Ajax: once, then restored after he appealed, then banned again. OpenAI has not publicly confirmed the reason. What the reporting connects it to is the way he gathered training data. He described using OpenAI's models to help produce and process examples for training his own model, and talked about "distilling its reasoning". Distillation means training a smaller model on the outputs of a larger one so it picks up some of the larger model's behavior.

OpenAI's terms of use forbid using its output to develop models that compete with OpenAI, and the company has been actively enforcing that rule. On July 1, 2026, it announced it had disrupted a distillation campaign by people linked to China's Moonshot AI. There is no reported connection between that case and PewDiePie, but it shows how seriously the company takes the pattern. For an ordinary user, the lesson is simple: if you ever fine-tune a model of your own, do not feed it bulk output from a commercial chatbot whose terms forbid it. Use open models, which is exactly what Ajax ended up being built on.

The bans also explain part of the appeal. A creator saying "the big AI company banned me for building my own" makes a stronger case for self-hosting than any spec sheet. That is a story about control: who decides what your assistant will do, and who can switch it off.

Odysseus and Ajax system requirements: what your PC needs

"odysseus ai system requirements" is one of the most common questions, and the answer depends on a choice the coverage rarely spells out: Odysseus itself is light. The model is what is heavy. The workspace is a web app that runs happily on a modest machine. The model it talks to decides whether you need a graphics card.

Odysseus can also connect to a model running somewhere else: another computer on your network, a rented GPU server, or a paid cloud API. So the true minimum is "a computer that can run Docker and a browser". Most people searching for PewDiePie's AI want the private, offline version, though, and that means running the model on the same machine. Here is what that takes for the Ajax base model, Qwen 3.5 9B, at the sizes Ollama offers.

A word on that "keep the context window modest" note. The model file is not the only thing that needs room on the graphics card. Every conversation also builds up a working memory, and the longer the conversation or document, the bigger it gets. On an 8 GB card the 6.6 GB model leaves a little over a gigabyte for that. Fine for email and short research tasks. Tight for feeding it a 60-page PDF. If replies suddenly crawl, part of the model has spilled from the graphics card into ordinary memory, and the fix is a shorter context or the next size of card. The honest laptop list for local models has the buying side of this if your current machine is not going to make it.

The rest of the machine matters less than people expect. Odysseus wants a few gigabytes of disk for its containers plus whatever models you download, at least 16 GB of system memory if a model runs alongside your browser and everything else, and a reasonably recent Windows 10 or 11, a mainstream Linux, or macOS on Apple Silicon. Ethan's laptop, an RTX 4060 with 8 GB of video memory and 16 GB of system memory, sits right in the sweet-spot row.

How to install Odysseus on Windows (Docker, step by step)

Odysseus offers two routes on Windows: Docker, which is the recommended one, and a native launcher that runs it straight from Python. Docker is easier to keep clean and easier to remove later, so start there. If containers are a new idea, the short version is that Docker runs Odysseus in a sealed box with everything it needs, so it cannot clutter the rest of your system.

Before you start: install the model runner

1. Download and install Ollama for Windows from ollama.com. It runs in the background and puts a small icon in the system tray.
2. Open PowerShell and download the Ajax base model: ollama pull qwen3.5:9b. That is the 6.6 GB file, so give it a few minutes.
3. Test it on its own first: ollama run qwen3.5:9b, ask it anything, and type /bye to leave. If this works, the hard part is done. If it does not, fix it here, before Odysseus is in the picture.

Install Odysseus with Docker Desktop

1. Install Docker Desktop for Windows and let it set up WSL 2 when it asks. Restart if prompted, then open Docker Desktop once so it finishes starting.
2. Install Git for Windows if you do not have it. Odysseus's Cookbook feature also uses it.
3. In PowerShell, go to a folder you are happy to keep it in, then run:

   git clone https://github.com/odysseus-dev/odysseus.git
   cd odysseus
   copy .env.example .env
   docker compose up -d --build

4. The first build takes several minutes. Check that everything is healthy with docker compose ps.
5. Get the first admin password, which Odysseus prints once during setup: docker compose logs odysseus. Look for the line with the temporary password.
6. Open http://localhost:7000 in your browser, sign in as the admin, and change the password in Settings straight away.
7. Add your model: in the model or provider settings, add an OpenAI-compatible endpoint with the address http://host.docker.internal:11434/v1 and choose qwen3.5:9b. The host.docker.internal part is how a program inside Docker reaches Ollama running on Windows itself. localhost would point back inside the box.

That is it. You now have PewDiePie's workspace running on your own machine, with the same model family Ajax uses, and nothing leaving your computer unless you connect a cloud service or switch on web search.

The native Windows route (no Docker)

If Docker will not run on your machine, Odysseus ships a one-command PowerShell launcher. It needs Python 3.11 and Git:

git clone https://github.com/odysseus-dev/odysseus.git
cd odysseus
powershell -ExecutionPolicy Bypass -File .\launch-windows.ps1

The -ExecutionPolicy Bypass part applies only to that one script run and does not change your system-wide setting. In native mode Odysseus runs on your machine directly, so the model address is the plain http://localhost:11434/v1. One limit: the vLLM and SGLang model servers need Linux, so on native Windows stick with Ollama, LM Studio or llama.cpp. The comparison of local model tools helps if you are not sure which runner suits you.

How to install Odysseus on Linux, Kali and a Mac

The Linux install is the same four commands as Windows, with one network detail that trips up almost everyone.

Linux and Kali with Docker

curl -fsSL https://ollama.com/install.sh | sh
ollama pull qwen3.5:9b
git clone https://github.com/odysseus-dev/odysseus.git
cd odysseus
cp .env.example .env
docker compose up -d --build
docker compose logs odysseus

On Kali and other Debian-based systems, install Docker first with sudo apt install -y docker.io plus the Compose package your release provides. If docker compose answers that "compose" is not a docker command, your system has the older standalone tool, and the same commands work as docker-compose with a hyphen. If you have an NVIDIA or AMD card, Odysseus includes separate compose files for GPU support, and scripts/check-docker-gpu.sh tells you whether Docker can see your card.

Now the detail. On Windows and Mac, Docker Desktop makes host.docker.internal point at your own machine automatically. On plain Linux it may not. And Ollama, by default, listens only on 127.0.0.1, which a container cannot reach. If Odysseus says it cannot reach the model, you have two choices. You can make Ollama listen more widely with OLLAMA_HOST=0.0.0.0:11434 ollama serve and point Odysseus at your machine's address on the Docker bridge, usually http://172.17.0.1:11434/v1. Or you can skip Docker for Odysseus entirely and use the native Linux install below, where plain localhost works. If you open Ollama to 0.0.0.0, make sure your firewall blocks port 11434 from outside your machine, because Ollama has no password of its own.

Linux without Docker

git clone https://github.com/odysseus-dev/odysseus.git
cd odysseus
python3 -m venv venv
source venv/bin/activate
pip install -r requirements.txt
python setup.py
python -m uvicorn app:app --host 127.0.0.1 --port 7000

Apple Silicon Mac

On a Mac, skip Docker. Docker on macOS cannot use the Mac's built-in GPU, so models would crawl on the CPU. Use the native launcher instead:

git clone https://github.com/odysseus-dev/odysseus.git
cd odysseus
./start-macos.sh

Then open http://127.0.0.1:7860, not 7000. The Mac launcher uses a different port because macOS's AirPlay Receiver already sits on port 7000, which is the single most common "it installed but the page won't load" report from Mac users.

One more note for all platforms: the repository's default branch is dev, which gets the newest changes first and occasionally breaks. If you want the steadier version, clone with git clone --branch main https://github.com/odysseus-dev/odysseus.git. Dependencies are not pinned, either, so two installs a few months apart may not behave identically. That is normal for a project moving this fast, and a good reason to note which version you installed.

Connect a local model: the Ajax base, running tonight

Odysseus does not care where its model comes from, as long as it speaks the OpenAI-style API that almost every local runner now offers. That gives you real choice.

Once a model is connected, start with something that shows off what the workspace adds over a plain chat window. Ask it to research a question across several web pages and write a short report, or open the Compare view and put the 9B model side by side with a smaller one on the same question. That second test is the spirit of the old council video, and it teaches you more about model quality in five minutes than any benchmark chart.

Ethan's first test was the one everyone does: he asked it who PewDiePie is. The 9B model gave a decent answer with a slightly wrong subscriber count. Then he switched on web search in Odysseus and asked again, and it came back with the current number and a source. That is the difference a workspace makes. The model is the brain, and the tools around it are what make it useful.

When Ajax drops: how to switch to it in ten minutes

Here is the plan for release day, so you are not reading a dozen launch-day articles at midnight.

1. Get it from the official source only. Watch PewDiePie's data site and the odysseus-dev GitHub organization for the link. If a download appears first on some random site or a re-upload account, wait for the official one. Model files can be swapped, and an agent model will be reading your email.
2. Read the license before you use it. Especially if you want to use it for work. Today nobody knows what it will be.
3. Pick the size that matches the requirements table above. If a 4-bit GGUF is offered, that is the 8 GB-card version. 8-bit or FP8 suits 12-16 GB cards, and BF16 suits 24 GB cards.
4. Load it in your runner. If it is published to the Ollama library, it is one ollama pull. If it comes as a GGUF file, LM Studio and llama.cpp load it directly, and for Ollama you write a two-line Modelfile pointing at the file, then run ollama create ajax -f Modelfile.
5. Switch the model in Odysseus and leave Qwen 3.5 9B installed for a week, so you can use Compare to check that Ajax actually does better on your own tasks before deleting the base.
6. Review the agent's permissions before you hand it your real inbox. The next section is the checklist.

The point of setting everything up in advance is that release day becomes a model swap, not a setup project. You have already fought Docker, found the right address for Ollama and learned how the workspace behaves. Ajax just slots in.

What can Odysseus do? The features, honestly

Odysseus is easier to understand as a set of rooms than as a feature list. Here is what each one does and how useful it is with a 9B model on a home machine.

Chat and agents. The main room. Conversations with memory, file uploads, custom instructions saved as presets, and agents that can use tools: web search, files, a shell, and MCP, the Model Context Protocol, which lets you plug in tools other people have built. Shell access is the powerful and dangerous one: an agent that can run commands can fix things and break things. A 9B model handles simple multi-step tasks well and stumbles on long chains, so keep agent jobs short and specific at first.

Deep Research. Give it a question and it searches, reads several sources and writes a report. This is where a small local model surprises people, because the hard part, finding and reading the pages, is done by tools, and the model only needs to summarize and connect. Check the sources it lists, the same as you would with any research assistant.

Compare. Blind side-by-side answers from different models, with an optional combined answer. It is the council idea made practical, and the best way to decide which model deserves your graphics card.

Documents, notes and tasks. A document editor with AI suggestions that handles Markdown, HTML and CSV, plus notes, to-do lists and reminders. Nothing here is unique on its own. The value is having it in the same place the model can see.

Email and calendar. An IMAP and SMTP email client with triage, tags, summaries and draft replies, and calendar sync through CalDAV. This is the room Ajax was trained for. It is also where setup takes longest, because you enter your own mail server settings, and where the safety section matters most.

Cookbook. A hardware-aware helper that looks at your machine and suggests which models will fit, then helps you download and serve them. It is genuinely useful for beginners, and a good second opinion beside the requirements table on this page.

Extras. An image gallery and image editor, themes, sessions, web search settings and two-factor sign-in. One thing to be clear on: the documented feature list does not include an image generator. Searches for "can odysseus ai generate images" are common, and the honest answer is that the workspace edits and organizes images, while making new ones would need a separate image model and tool connected to it.

Odysseus vs ChatGPT, Ollama, Open WebUI, Hermes Agent and OpenClaw

Most comparison searches about PewDiePie's AI are really asking "do I need this, or do I already have something that does it?" Here is the plain answer for each.

On "is PewDiePie's AI better than ChatGPT": no, and nobody who has run a 9B model believes otherwise. A frontier cloud model knows more, reasons further and writes better. The case for Odysseus is everything else: your email, notes and research never leave your machine, nobody can ban your account or change the rules on you, and once it is set up, it costs only the electricity. If you are weighing the subscription side of that trade, our honest look at whether ChatGPT Plus is worth it runs the numbers. Plenty of people end up using both: a cloud model for the hardest questions, and a local workspace for anything private.

Running Odysseus safely: treat it like an admin console

The project's own documentation says to treat Odysseus "like an admin console", and it means it. Out of the box it can run shell commands, manage files, hold API keys and, once you connect them, read your email and calendar. That is a lot of power for a browser tab. These settings and habits keep it under control.

1. Keep it on localhost. Odysseus binds to 127.0.0.1 by default, which means only your own computer can open it. Do not change APP_BIND to 0.0.0.0 unless you know exactly who can reach that port.
2. Keep sign-in on. Leave AUTH_ENABLED=true and LOCALHOST_BYPASS=false, change the first admin password immediately, and switch on two-factor sign-in.
3. Never expose the raw port to the internet. If you want it from your phone, use a private network layer such as Tailscale, a VPN or Cloudflare Access, with HTTPS in front. A port-forward from your router is how home servers get found by strangers.
4. Do not expose Ollama either. If you made Ollama listen on 0.0.0.0 for Docker on Linux, firewall port 11434. Ollama has no password, and an open one lets anyone use your graphics card.
5. Give the agent the smallest inbox you can. Start with a secondary email account, not the one your bank writes to. Let it draft replies, not send them, until you trust what it does.
6. Be careful with shell access. An agent that can run commands can be talked into running the wrong ones by a web page it reads. If you want an agent with real system access, run it inside a sandbox; NVIDIA's OpenShell sandbox is one way to fence an agent in on Linux.
7. Pin a version once it works. The project suggests fixed image tags over :latest for anything you rely on, so an overnight update does not change behavior under you.

This list matters more with Ajax than with a standard model, for the reason in the uncensored section: the refusal training that sometimes makes a model hesitate has been deliberately removed. Most of the time that is exactly what users want. When the model is reading untrusted web pages and emails while holding your accounts, the guardrails have to live in the setup instead.

Odysseus not working? Common errors and fixes

Nearly every "odysseus ai not working" report comes down to one of six problems.

Two habits prevent most of these. Always test the model on its own before blaming Odysseus, because if ollama run qwen3.5:9b does not answer, nothing built on top of it will. And when you ask for help on the project's issue tracker or Reddit, paste the last hundred or so lines of the logs. "It doesn't work" gets sympathy. A log gets an answer.

Is PewDiePie's AI legit? Official links and fake downloads

Yes, the project is real and open for anyone to inspect: the code is public, the license is a standard open-source one, and tens of thousands of developers have looked at it. The risk is not Odysseus. It is the ecosystem that springs up around any famous name.

Within days of a celebrity tech launch, websites appear with the name in the address, offering downloads, "online versions" and setup guides. Some are harmless fan pages. Some are advertising farms. A few bundle something you would never knowingly install. For PewDiePie's AI, the official places are short: the odysseus-dev organization on GitHub for the workspace, and PewDiePie's own data site for Ajax announcements. Anything else is third-party, however official it looks.

Three quick checks keep you safe. If a site offers an "Odysseus AI online" login, remember that Odysseus is self-hosted: the real one runs on your machine, so a hosted login is someone else's service. If a site offers an Ajax download before PewDiePie announces one, it is not Ajax. And if anything asks you to disable your antivirus to install it, close the tab. The real install is four commands and a browser.

Does PewDiePie's AI use water? The environment question

"does pewdiepie's ai use water" is a surprisingly common search, and it is a fair question, because the water story around AI is mostly about data centers. Large cloud facilities cool their servers with systems that evaporate water, and that is where the headline numbers about water per conversation come from.

A model running on your own PC does not use evaporative cooling. Your graphics card is cooled by fans, or by a sealed liquid loop that does not consume water. So Odysseus with a local model uses no water directly. What it uses is electricity, and power plants can use water, so the honest answer is "only indirectly, through your power bill, like gaming does".

The electricity is easier to picture than people expect. A laptop RTX 4060 draws somewhere around 100 watts at full load, and a 9B model answering a question keeps it busy for a few seconds. That is roughly the energy of leaving a light bulb on for a minute or two. A day of heavy use costs about what an evening of gaming does. PewDiePie's ten-card rig is another matter, and nobody needs one to run Ajax. That is rather the point of building a 9B model.

PewDiePie AI, Odysseus and Ajax: frequently asked questions

Ethan left the shop that afternoon with Odysseus on his laptop, Qwen 3.5 9B answering through it, and a note on his phone from Jake: "When Ajax is out, official link only, then ollama pull and switch it in Settings. Do not give it your main email." He set up a spare account that evening, let the workspace sort a week of junk mail into folders, and sent Jake a screenshot with a single word: "Mine." That is the whole appeal in one word. Not that it beats the biggest cloud models, because it does not, but that it lives on your machine, answers to you, and keeps working whatever any company decides. Ajax will make it a little smarter at its job. The setup you build today is what makes it yours.

Revision note. Written October 2, 2026, the day Ajax was announced; the release-day section gets a dated update when it ships. If you set this up for someone you love, make sure their spare inbox is the one it reads.