Alright, so if you’ve read the title and thought you were having a stroke, you might not be the target audience for this post. If, however, you thought “more tokens/s?”, you should keep on reading.

If you remember my review of the new Lenovo X1 Carbon Gen 14 Aura with Intel Panther Lake Core Ultra X7 368H vPro from a while ago, you might recall that I had tested its local “AI” performance using Ollama, via Vulkan, which obviously didn’t perform particularly well across various models. The proper way to run local LLMs is to use Intel’s official oneAPI framework and compile llama.cpp with SYCL enabled, which will greatly increase performance.

This post is a very brief write-up of how that can be done and primarily serves as yet another documentation for future-me.

Gentoo

I’m assuming that you’re already running your Gentoo system with the xe driver and VIDEO_CARDS="intel". First, install the necessary dependencies:

emerge -av dev-libs/intel-compute-runtime dev-libs/level-zero \
           dev-util/intel-graphics-compiler media-libs/gmmlib \
           dev-util/clinfo

You’ll likely have to unmask a couple of those packages, like clinfo, but that’s fine. When this is done, add your user to the video and render groups if you haven’t already and log back in. Test that the GPU is being recognized by clinfo -l:

Platform #0: Intel(R) OpenCL
 `-- Device #0: Intel(R) Core(TM) Ultra X7 368H
Platform #1: rusticl
Platform #2: Intel(R) OpenCL Graphics
 `-- Device #0: Intel(R) Arc(TM) B390 GPU

oneAPI

Next, go and get the “Intel Deep Learning Essentials” package and use ./intel-deep-learning-essentials-2026.1.2.25_offline.sh -a --cli to install it as a user. I chose the installation path ~/projects/intel/oneapi, but you’re free to install it anywhere you please.

Note: Even if you deselect the telemetry option, the installer will try to contact Intel’s servers post-installation, so make sure your OpenSnitch blocks all requests from that process towards the interwebs.

llama.cpp

While the FireBurn overlay has sci-misc/llama-cpp, it does not yet allow you to specify sycl as a USE flag, and the official Gentoo repository has no llama.cpp ebuild at all. Hence we clone the project’s Git repository and compile it manually:

git clone https://github.com/ggml-org/llama.cpp
cd llama.cpp
source ~/projects/intel/oneapi/setvars.sh
cmake -B build -DGGML_SYCL=ON -DGGML_SYCL_F16=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx -DCMAKE_BUILD_TYPE=Release
cmake --build build --config Release -j

Run

Once llama.cpp finishes building, you can start it and have it download a model, e.g., Mistral or Qwen3:

source ~/projects/intel/oneapi/setvars.sh

./build/bin/llama-server -hf bartowski/Mistral-7B-Instruct-v0.3-GGUF:Q4_K_M -ngl 99 -sm none -mg 0 -c 8192 --host 127.0.0.1 --port 8080

or:

./build/bin/llama-server -hf Qwen/Qwen3-4B-GGUF:Q4_K_M --jinja -ngl 99 --temp 0.6 --top-k 20 --top-p 0.95 --min-p 0 --presence-penalty 1.5 -c 40960 -n 32768 --no-context-shift --host 127.0.0.1 --port 8080

Open a browser at http://127.0.0.1:8080, and you can try the model right away.

On my Lenovo, I managed to increase the tokens/second by roughly 60% compared to what Ollama (via Vulkan) was able to achieve. For example, the Mistral model went from approximately 13.88 tokens/s to 22.36 tokens/s.