
Core Announcement and Operational Scope
The ggml-org/llama.cpp project has published pre-release build b12502 on GitHub, introducing a targeted CUDA execution update for Mixture-of-Experts (MoE) workloads alongside a broad automated binary release spanning desktop, mobile, and server architectures. Triggered by github-actions on November 18, 2026 at 08:09 UTC with a published timestamp of 2026-11-18T08:14:05+00:00, the release packages commit 398f00d as a pre-release build containing 40 downloadable assets and a UI package.
Rather than introducing broad API changes, b12502 centers on a single documented pull request—#32504, titled "CUDA: use MMQ for MoE on V100"—while maintaining automated multi-backend artifact generation for local and server-side model execution. For teams tracking upstream llama.cpp snapshots, the release establishes a traceable checkpoint linking source commit, cryptographic signature, build attestation, and precompiled runtimes.
CUDA MMQ MoE Execution on V100 and Provenance
At the technical core of b12502 is pull request #32504, which routes CUDA Mixture-of-Experts workloads to MMQ kernels on NVIDIA V100 hardware. Within the llama.cpp CUDA backend, MMQ handles matrix multiplication on quantized weights directly, making the V100 routing change relevant to operators running quantized MoE models on Volta-generation datacenter GPUs. Because the release page itself does not publish tokens-per-second figures, latency measurements, or memory benchmarks for #32504, real-world performance effects on V100 nodes remain unquantified in the release notes and must be measured against specific models and quantization formats.
The release metadata also records explicit supply-chain verification details. Commit 398f00d was created on GitHub.com and signed with GitHub's verified signature using GPG key ID B5690EEEBB952194. In addition, the release notes reference a GitHub build attestation record (attestations/69023761), giving integrators verifiable provenance linking the published archives back to the automated CI build.
Cross-Platform Build Matrix and Hardware Targets
Despite its focused code changelog, b12502 distributes an extensive set of prebuilt binaries across CPU, GPU, and NPU backends. On Linux, Ubuntu targets include x64 and arm64 CPU builds, IBM mainframe builds for Ubuntu 24.04 s390x (CPU) and Ubuntu 26.04 s390x (both CPU and zDNN), Vulkan builds for x64 and arm64, OpenVINO on x64, SYCL FP32 and SYCL FP16 on x64, and ROCm 10.0 on x64. Linux CUDA packages span Ubuntu x64 with CUDA 12.8 libraries as well as Ubuntu x64 and Ubuntu arm64 builds using CUDA 13.5 libraries, while Linux arm64 Snapdragon builds target CPU, Adreno GPU, and Hexagon NPU execution.
Windows coverage mirrors much of this heterogeneity, providing x64 and arm64 CPU binaries alongside Windows arm64 OpenCL Adreno, x64 and arm64 Vulkan, x64 OpenVINO, x64 SYCL, and x64 ROCm 10.0 builds. Windows CUDA releases cover x64 with CUDA 12.4 DLLs and both x64 and arm64 builds shipping CUDA 13.5 DLLs. In the Apple and mobile ecosystems, the matrix lists macOS Apple Silicon (arm64), macOS Intel (x64), an iOS XCFramework, Android arm64 (CPU), and Android arm64 for Snapdragon (CPU, Adreno GPU, and Hexagon NPU), alongside packaged UI assets and a project website reference at llama.app.
Developer and Deployment Workflow Impact
For infrastructure engineers and open-source AI developers, the immediate benefit of b12502 is lower packaging friction across mixed hardware fleets. Teams operating older V100 servers alongside newer CUDA 12 or CUDA 13 workstations can pull matching prebuilt libraries—such as CUDA 12.8 on Ubuntu x64, CUDA 12.4 DLLs on Windows x64, or CUDA 13.5 on Linux and Windows x64 and arm64—without compiling custom toolchains from source. Likewise, the simultaneous publication of s390x zDNN, ROCm 10.0, SYCL FP16/FP32, OpenVINO, Vulkan, and Snapdragon Hexagon NPU targets allows toolmakers to test cross-vendor runtime parity from a single tagged commit.
Having a signed commit and a linked GitHub attestation record also supports automated provenance checks for engineering teams that gate pre-release binary ingestion before staging updates into internal inference services or local developer tools.
Caveats, Disabled Targets, and Verification Boundaries
Several caveats apply before deploying b12502 in production. First, the release is explicitly tagged as a pre-release snapshot generated by automated workflows rather than a stabilized milestone. Second, two target entries are marked DISABLED in the release notes: the macOS Apple Silicon (arm64, KleidiAI enabled) build and the openEuler section covering x86 and aarch64 builds for 310p and 910b ACL Graph hardware. Teams depending on KleidiAI on macOS or openEuler Ascend packages should check asset availability directly before updating download scripts.
Finally, the release page excerpt does not state repository licensing terms, driver prerequisites beyond listed CUDA library versions, or independent benchmark validation for pull request #32504. Operators should verify the upstream repository license, review the pull request diff, and run controlled local validation on their target MoE models and hardware accelerators prior to rollout.
EXPLORE THE SOURCE
Projects in this story
GitHub repositories
- ggml-org/llama.cpp ↗View project details
Automated Vertex AI (gemini-4-hydrogen) primary-source report; reviewed 2026-11-18.



