PewDiePie has revealed Ajax, a local AI model he’s building as a personal assistant for his Odysseus workspace. It’s designed to run on your own hardware, work with your everyday tools, and refuse fewer requests.
Ajax is a fine-tuned version of Qwen3.5-9B, a model with around 9 billion parameters. PewDiePie is taking an existing model and training it specifically for the things he wants his assistant to do.
It isn’t out yet. As of October 5, 2026, the official Ajax page still says it’s coming soon. PewDiePie dropped the countdown and says he’ll release it when it’s ready, so for now, all we have is what Ajax promises.
Ajax Is Being Trained for Odysseus
PewDiePie’s development thread describes plans for a small model that understands Odysseus tools, makes valid tool calls, handles the results, and answers without dragging every simple task into a long reasoning session.
If you ask it to create a reminder, it should choose the reminder tool, send the correct information, and tell you what happened. If a tool fails, it should recognize the problem and handle it properly.
That sounds like a simple task, but anyone who has tried getting a smaller local model to use tools knows it does not always go smoothly. PewDiePie lists misunderstood tool descriptions, invalid calls, and models struggling with the full agent prompt as reasons for building a dedicated fine-tune.
A model that reliably finishes a small job would be really useful, especially when the alternative is watching it explain how it plans to finish the job.
Odysseus Connects Ajax to Your Tools
Odysseus is PewDiePie’s open-source, self-hosted AI workspace. You run it on your own computer or server and use it to bring chats, documents, email, notes, and calendars into one place. It supports local models and external model APIs, along with tools the assistant can use to get things done.
So in theory, Ajax could use Odysseus’s email and notes tools to summarize your recent emails and jot down notes about something mentioned in one of them. And if you want to see how tool calling works, you can read our tool calling in Ollama article.
How Ajax Reduces Refusals
The official page says Ajax’s refusal behavior has been ablated. That means the model has been modified to reduce its tendency to reject requests.
Refusal ablation changes behavior inside the model itself. Tools such as Heretic modify model weights associated with refusals while trying to preserve the model’s other abilities. Changing this behavior can also affect how well the model answers.
Agreeing to a request doesn’t mean getting it right, though. Ajax could accept your reminder request and then set the wrong date. What matters is which requests it still turns down and how reliably it handles the ones it takes on.
Ajax’s Planned 4-Bit Quantized Release
PewDiePie’s development thread lists a Q4 quantized release as a target for broader local use. It also lists GGUF files for llama.cpp and support for vLLM.
A smaller quantized build should make Ajax easier to run on personal devices, but parameter count alone won’t tell you if it fits your machine. As we explained in our quantization guide, lowering the precision of the weights cuts memory use, though you still need room for context and everything else inference needs.
For now, we need the actual release formats and some real testing before giving a specific RAM or VRAM requirement.
Conclusion
The real test will be how often Ajax completes everyday tasks correctly. If it can do that reliably on ordinary local hardware, it could become an assistant people actually keep running.
Sources
PewDiePie’s Odysseus fine-tuning discussion
Heretic refusal ablation documentation
Official Ajax page
