A-TIER skill
Abliterated Qwen 3.5 9B — run PewDiePie-style Ajax locally with llama.cpp
100k ft — What is this?
Abliterated Qwen 3.5 9B — run PewDiePie-style Ajax locally with llama.cpp
50k ft — Why care?
Walkthrough of running an abliterated Qwen 3.5 9B GGUF (the base of PewDiePie's Ajax model) locally: download from Hugging Face, install llama.cpp via Homebrew, start llama-server with context size and port, then point the DeepSeek harness at the custom endpoint. Also mentions PewDiePie's Odysseus agent harness.
10k ft — Key takeaways
• A 16-bit 9B model (~17 GB) runs on a 32 GB MacBook Pro
• Abliterated models remove refusals more thoroughly than plain uncensored fine-tunes
• Any agent harness with a custom OpenAI-compatible endpoint can drive it
• brew install llama.cpp then llama-server -m <model> --ctx-size --port 8080
• link: https://brew.sh
5k ft — Tools & techniques
DeepSeek, DeepSeek Harness, Homebrew, Hugging Face, Llama, Odysseus, OpenAI, Qwen, llama.cpp
AI agents, GGUF quantization, abliteration, local inference
1k ft — Deep-dive / try it
Let's try out the new Ajax model Pewdiepie created...Kinda 🔥🙌
Pewdiepie is an online streamer and gamer, and earlier this year created his own Agent Harness called Odysseus to run his LLM models locally. Odysseus agent became one of the top popular and downloaded repo on github.
He went on to create an abliterated model called Ajax by using Qwen 3,5 9B as the base model. He was subsequently banned from OpenAI for training his LLM model.
Follow @codrkai
#pewdiepie #ajax #qwen #qwenai #deepseek
Transcript excerpt
So PewDiePie who created Odysseus created a new model that is unrestricted and obliterated. It's called Ajax. Let's try it out. Go to hugandface.co and type in quen 3.5 9b obliterated. This is the same model that PewDiePie based his off of. There are uncensored models and obliterated models. The obliterated model is generally better for removing restrictions. So we're going to use that one. I'm going to select this obliterated guff model and I'm going to download the full 16 bit 17 gigabyte version because I have a 32 gigabyte MacBook Pro that can handle this. So I created a models folder in the root of my computer to store my models. To run the model locally this time I'm going to use llama.cpp as the server. This has been requested by some of you. To install it I'm using homebrew. Go to brew.sh and copy the curl command. Then open up terminal, paste it in and run it to install. Then type brew install llama.cpp and it's that easy. GG freaking ease. Now let's run the server with a model by typing llama-server-m. Paste in the model path dash dash ctx-size. I need to give it this context size for my MacBook Pro. Give it a port 8080 and then these command help it run eventually on my MacBook Pro. Ginger and parallel 2. Once the server is up and running I'm now going to use deep seek harness as my agent for this example. I have a video on deep seek harness if you want to watch that one. Now I need to go in my settings and set the model provider. To set my custom endpoint I'm…