llama.cpp b9628 ships SYCL support, broadening GPU access beyond CUDA

The latest release of llama.cpp, tagged b9628, ships one notable addition: SYCL support. The release notes on GitHub list a single change — “add sycl to check-release (#24583)” — but the downstream assets tell a bigger story. For the first time, the project distributes prebuilt binaries for SYCL FP32 and SYCL FP16 on Ubuntu x64, plus a SYCL build for Windows x64.

SYCL is an open, royalty-free programming model for heterogeneous computing. It allows the same code to run on GPUs from Intel, AMD, and NVIDIA, as well as CPUs and FPGAs. For llama.cpp, a project that has long been the default way to run large language models on consumer hardware, SYCL support means one thing: Intel GPU owners can now run local inference without a CUDA workaround.

This is not a headline-grabbing release. There is no new model, no performance benchmark, no architectural breakthrough. But it is the kind of release that quietly reshapes the AI inference landscape.

The b9628 asset list is a catalog of hardware diversity. Ubuntu builds ship for CPU, Vulkan, ROCm 7.2, OpenVINO 2026.0, and now SYCL. Windows builds include CUDA 12, CUDA 13, Vulkan, HIP, and SYCL. There is even an s390x build for IBM mainframes. The macOS Apple Silicon builds include a KleidiAI variant, though it is currently disabled pending a fix. The project now targets more hardware backends than any commercial inference platform.

What makes this significant is the timing. The AI hardware market is in the middle of a slow but real diversification. NVIDIA still dominates training and large-scale inference, but the narrative around GPU lock-in is fraying. AMD’s MI300 series has found buyers. Intel’s Gaudi line is shipping. Cloud providers are building custom silicon. And on the desktop, Intel’s Arc GPUs and integrated graphics represent a massive installed base that has been largely excluded from local AI inference.

SYCL is the bridge. Intel has been pushing SYCL as the standard for heterogeneous computing on its hardware for years, with limited success in the AI space. Most AI frameworks default to CUDA. PyTorch and TensorFlow have SYCL backends, but they are secondary citizens. llama.cpp, by contrast, has always been backend-agnostic by design. The ggml tensor library at its core was built to run on anything. Adding SYCL is not a pivot. It is a natural extension of the project’s original philosophy.

The practical effect is modest but real. An Intel Arc A770 owner, or someone with a laptop running Intel Iris Xe graphics, can now download a single binary and run Llama 3 or Mistral locally. No CUDA toolkit. No WSL2 workaround. No Docker container with NVIDIA Container Toolkit. Just a download and a command.

That matters for the AI culture that llama.cpp has fostered. The project is the backbone of a whole ecosystem of local AI tools: text generation UIs like Ollama and LM Studio, agent frameworks, RAG pipelines, and privacy-focused chatbots. Every time llama.cpp adds a backend, that ecosystem expands to new hardware. SYCL brings Intel GPU users into the fold.

The release also signals something about the project’s maturity. The b9628 tag includes a “check-release” CI step for SYCL, meaning the build is tested and maintained. This is not a proof-of-concept. It is a production-grade integration that the maintainers intend to keep working. For enterprise users who need to run inference on Intel GPU clusters — and there are such deployments, particularly in Europe and China where Intel hardware is more common — this is a meaningful step.

There are limits. SYCL performance on Intel GPUs is not going to match CUDA on an RTX 4090. The Intel GPU software stack is still maturing. And the SYCL standard itself has fragmentation between Intel’s oneAPI implementation and the open-source triSYCL project. But the baseline is now there, and it will improve.

The broader take is about the shape of AI infrastructure. For two years, the conversation has been about NVIDIA’s dominance and the difficulty of running models on anything else. llama.cpp b9628 is a small, concrete counterexample. It shows that open-source software can be the vehicle for hardware diversity, one backend at a time.

The next release will add something else. Maybe Flash Attention for Vulkan. Maybe a new quantization format. Maybe a fix for the disabled KleidiAI build. The project never stops. But b9628 will be the one where Intel GPU owners could finally join the party. That is the kind of change that does not make headlines. It just makes the technology work for more people.