How to run Ajax locally (guide skeleton)
Ajax is not released - Not released - so this page is a skeleton. It separates what you can actually do today from the steps that only exist once the weights ship. Nothing below invents commands we can't verify.
What "running Ajax locally" actually means
People searching this usually mean one of two different things, and they need different setups. The first isrunning the weights: a 9-billion-parameter model file loaded into a local runtime like Ollama or LM Studio, chatting with it like any other open model. That path needs nothing PewDiePie-specific - just the file and enough VRAM.
The second is running Ajax as intended: the always-on private agent inside Odysseus, the self-hosted workspace it was trained for. That's the version that reads your email, searches the web and manages your calendar without data leaving your machine. It needs Odysseus running plus the model file - the model alone won't give you the agent loop.
This page covers both. Path 1 is the intended one; Paths 2 and 3 are for people who just want the raw model.
Path 1: Inside Odysseus (the intended way)
Ajax is "trained for Odysseus to be your always on agent" - Odysseus is the workspace that gives the model tools: search, browsing, email, calendar, notes. If you want the full always-on agent experience, this is the path.
What works today:
- Clone the workspace:
git clone https://github.com/odysseus-dev/odysseus - Read the docs at odysseus-dev.github.io/odysseus
- The repo ships a
docker/folder plusmcp_servers/,integrations/andconfig/searxng/- the services that make Odysseus an agent harness rather than a chat UI. Odysseus already runs other local models, so you can set the whole thing up now.
Filled in on release day:
- Pulling the Ajax weights from the official link and placing them where Odysseus expects local models
- Any Ajax-specific config Odysseus adds for the model
- Verified setup steps once we can test them
Path 2: Ollama
If the release includes a format Ollama can load (GGUF is the usual route for 9B models), running Ajax should be a one-liner - but no official Ollama model tag exists and we won't guess one. What we can say today:
- Install Ollama from ollama.com as usual.
- If a GGUF ships without an Ollama tag, the generic path is a
ModelfilewithFROM ./file.ggufthenollama create- standard Ollama usage, not Ajax-specific.
Release-day steps (exact tag/Modelfile, quant choice) go here once files exist.
Path 3: LM Studio
LM Studio loads GGUF files through a GUI, which is the friendliest option if you'd rather not touch a terminal. Today: install LM Studio and confirm your machine matches the requirements estimates. On release day this section will note which quant fits which GPU and any chat-template quirks the fine-tune introduces.
The hardware question
Whether you can run Ajax locally at all comes down to GPU memory, and the estimates have a decent spread. Full precision (bf16/fp16) on the 9B base is reported at roughly 22 GB - RTX 3090/4090 territory. An 8-bit build is about 11 GB, which squeezes onto a 12 GB card like the 3060 12G or a 4070. Our derived estimate for a 4-bit GGUF-style build is ~5–6 GB, which opens the door to 8 GB cards and 16 GB unified-memory Macs. All three tiers are estimates - nothing is confirmed until files ship, and context length pushes these numbers up. Details and the working table live on the requirements page.
Questions people actually ask
- Do I need Odysseus to run Ajax? No - if an open format ships (GGUF etc.), any runtime can load the weights. Odysseus is only needed for the agent features it was trained for.
- Will there be an official Ollama tag? Unknown. If not, a
Modelfile+ollama createworks for GGUF files, or community mirrors will appear (unofficial - verify before trusting). - Can I run it on a Mac? Plausibly yes at 4-bit on 16 GB unified memory - our estimate, not a promise.
- When can I actually do this? Whenever the official page gains real files. The badge on the status page flips automatically.
What happens here on release day
The moment the tracker detects a real download link on the official page or an official Hugging Face repo, this page flips: the skeleton text is replaced by verified commands - the exact HF repo, huggingface-cliand ollama run lines, and LM Studio search hints - pulled from the detected release data. It also gains a fake-repo warning, because re-uploads with the same name will appear within hours.
Honest summary
The only step that exists right now is setting up Odysseus - everything model-side waits on the official page flipping from "Coming soon" to actual files. Subscribe below and this page becomes a real guide the day that happens.
Get pinged the day Ajax ships
One email when the status on this page changes to released. Nothing else.
Stored only to send the release notification. See privacy.
Related: Ajax vs Odysseus, download status, news.