Uploaded June 2026 | Updated September 2026, 2 weeks ago
PewDiePie's Project Odysseus promises a slick all-in-one front end for running local LLMs — but what does setup actually look like if you're not running four RTX 6000 Pros? We find out live.
I'm running an Intel Arc Pro B70 (32GB VRAM, ~$1,000) and taking Odysseus from a fresh clone all the way to a working local AI assistant — hitting every GPU detection failure, broken dependency, and syscl build error along the way.
By the end of the stream we have Qwen 3 35B-A3B (MoE, Q5) running locally through llama.cpp with SYCL backend, wired into Odysseus, and doing real deep-research tasks — including drafting a port plan for DankMaterialShell from Wayland to X11.
▶ What we cover:
- Hardware reality check — what you actually need vs. what Pewds used (~$40K in RTX 6000 Pros)
- Native vs. Docker install (and why Intel GPU users need native)
- Building llama.cpp from source with Intel SYCL/oneAPI backend
- GPU detection bug in Odysseus + open PR walkthrough
- Picking and downloading the right Qwen 3 model for 32GB VRAM
- Context sizing math — hitting 128K context on a Q5 quant
- Writing a systemd service to auto-launch llama-server
- SearXNG integration + Context7 MCP setup
- Deep research demo: porting a Wayland shell to X11
- Wayland vs. X11 window manager rant (RIP MangoWM)
⚡ For most people: if you have a high-end NVIDIA card, just use the Docker install and you'll be up in ~30 min. Intel/AMD users, buckle up.
📖 The Linux Desktop Guide (physical book): christitus.com
🛠 WinUtil: github.com/ChrisTitusTech/winutil
🐧 Linutil: github.com/ChrisTitusTech/linutil
🔴 Stream schedule: Mon / Tue / Thu
💬 Community forum: forum.christitus.com
#Linux #LocalAI #LLM #ProjectOdysseus #IntelArc #llamacpp #Qwen
00:00 Intro & hardware reality
01:54 Intel B70 overview
03:27 Project Odysseus overview
06:54 Native vs. Docker decision
08:50 PewDiePie's $40K GPU build
11:31 Cloning & installing Odysseus
18:30 Install complete
22:14 GPU not detected
26:34 Intel support bug & PR
31:33 Building llama.cpp from source
37:44 Reinstalling oneAPI toolkit
42:45 SYCL AUR package attempt
53:20 Build finally succeeds
58:08 What we learned so far
1:05:14 DistroBox attempt fails
1:12:02 Back to native approach
1:17:18 Full clean rebuild
1:28:46 Past-Chris saves the day
1:40:07 Llama server running
1:51:13 Odysseus connected
1:53:31 First AI test
1:54:40 Deep research demo
2:06:18 PewDiePie 200 IQ moment
2:09:27 Downloading Qwen 3 35B
2:26:47 128K context confirmed
2:29:25 Qwen 3 live in Odysseus
2:30:50 Systemd service with AI
2:56:40 Full system working
2:57:44 SearXNG integration
2:59:32 MCP server setup
3:08:28 Wrap-up & next stream
►► Digital Downloads ➜ cttstore.com
►► Patreon ➜ patreon.com/christitustech
►► Twitch ➜ twitch.tv/christitustech
►► Music by: CreatorMix.com
PewDiePie's Project Odysseus promises a slick all-in-one front end for running local LLMs — but what does setup actually look like if you're not running four RTX 6000 Pros? We find out live.
I'm running an Intel Arc Pro B70 (32GB VRAM, ~$1,000) and taking Odysseus from a fresh clone all the way to a working local AI assistant — hitting every GPU detection failure, broken dependency, and syscl build error along the way.
By the end of the stream we have Qwen 3 35B-A3B (MoE, Q5) running locally through llama.cpp with SYCL backend, wired into Odysseus, and doing real deep-research tasks — including drafting a port plan for DankMaterialShell from Wayland to X11.
▶ What we cover:
- Hardware reality check — what you actually need vs. what Pewds used (~$40K in RTX 6000 Pros)
- Native vs. Docker install (and why Intel GPU users need native)
- Building llama.cpp from source with Intel SYCL/oneAPI backend
- GPU detection bug in Odysseus + open PR walkthrough
- Picking and downloading the right Qwen 3 model for 32GB VRAM
- Context sizing math — hitting 128K context on a Q5 quant
- Writing a systemd service to auto-launch llama-server
- SearXNG integration + Context7 MCP setup
- Deep research demo: porting a Wayland shell to X11
- Wayland vs. X11 window manager rant (RIP MangoWM)
⚡ For most people: if you have a high-end NVIDIA card, just use the Docker install and you'll be up in ~30 min. Intel/AMD users, buckle up.
📖 The Linux Desktop Guide (physical book): christitus.com
🛠 WinUtil: github.com/ChrisTitusTech/winutil
🐧 Linutil: github.com/ChrisTitusTech/linutil
🔴 Stream schedule: Mon / Tue / Thu
💬 Community forum: forum.christitus.com
#Linux #LocalAI #LLM #ProjectOdysseus #IntelArc #llamacpp #Qwen
00:00 Intro & hardware reality
01:54 Intel B70 overview
03:27 Project Odysseus overview
06:54 Native vs. Docker decision
08:50 PewDiePie's $40K GPU build
11:31 Cloning & installing Odysseus
18:30 Install complete
22:14 GPU not detected
26:34 Intel support bug & PR
31:33 Building llama.cpp from source
37:44 Reinstalling oneAPI toolkit
42:45 SYCL AUR package attempt
53:20 Build finally succeeds
58:08 What we learned so far
1:05:14 DistroBox attempt fails
1:12:02 Back to native approach
1:17:18 Full clean rebuild
1:28:46 Past-Chris saves the day
1:40:07 Llama server running
1:51:13 Odysseus connected
1:53:31 First AI test
1:54:40 Deep research demo
2:06:18 PewDiePie 200 IQ moment
2:09:27 Downloading Qwen 3 35B
2:26:47 128K context confirmed
2:29:25 Qwen 3 live in Odysseus
2:30:50 Systemd service with AI
2:56:40 Full system working
2:57:44 SearXNG integration
2:59:32 MCP server setup
3:08:28 Wrap-up & next stream
►► Digital Downloads ➜ cttstore.com
►► Patreon ➜ patreon.com/christitustech
►► Twitch ➜ twitch.tv/christitustech
►► Music by: CreatorMix.com
