
Core Release Scope and Security Fix Overview
The open-source inference engine llama.cpp released tag b12556 on November 19, 2026, shipping a targeted memory-safety fix alongside an expansive set of prebuilt platform binaries. Tagged as a pre-release at 23:12 UTC from commit 50518b6 (and indexed in the release feed at 23:17:27 UTC), the build was generated by github-actions and signed with GitHub's verified GPG key B5690EEEBB952194, backed by public build attestations and 40 downloadable assets.
The sole code change documented in the b12556 changelog is pull request #32601, titled model : hrm : do not compute slot count as uint32_t. By eliminating a 32-bit unsigned integer wraparound during Hierarchical Reasoning Model (HRM) hyperparameter validation, the update prevents malformed or adversarial model metadata from bypassing bounds checks and triggering an out-of-bounds write during tensor initialization. For teams running local models sourced from public hubs, hardening the model loader directly protects host processes against corrupted headers before inference begins.
Memory Safety Mechanics and Cross-Platform Build Specifications
According to the release notes for #32601, the flaw stemmed from a type mismatch between validation in load_arch_hparams and iteration in load_arch_tensors. Specifically, load_arch_hparams multiplied three hyperparameter fields—n_hrm_layers_per_stack, n_hrm_h_cycles, and (n_hrm_l_cycles + 1)—into a uint32_t and then asserted that the resulting product equaled block_count. Because 32-bit unsigned multiplication wraps modulo 2^32, any combination of factors whose true mathematical product equals block_count + k*2^32 passes the equality assertion. The changelog illustrates this behavior with the concrete example 1073741825 * 2 * 2 = 2^32 + 4, which wraps to 4 and evaluates as equal to block_count = 4.
Crucially, load_arch_tensors subsequently iterates over those exact same three hyperparameter values using signed int arithmetic rather than the wrapped uint32_t representation. As a result, the tensor loader attempts to populate more slots than allocated, writing past the end of the block_count-sized layers vector. Removing the uint32_t slot-count calculation aligns validation with actual loop bounds and closes that out-of-bounds memory write path.
Alongside the loader fix, b12556 publishes binaries across five major operating system families and multiple accelerator toolchains:
- macOS and iOS: macOS Apple Silicon (arm64), macOS Intel (x64), and iOS XCFramework, while macOS Apple Silicon (arm64, KleidiAI enabled) is marked DISABLED.
- Linux: Ubuntu x64 and arm64 (CPU and Vulkan), Ubuntu 24.04 and 26.04 s390x (CPU, plus zDNN on Ubuntu 26.04), Ubuntu x64 with CUDA 12 (CUDA 12.8 libraries), Ubuntu x64 and arm64 with CUDA 13 (CUDA 13.5 libraries), Ubuntu x64 with ROCm 10.0, OpenVINO, and SYCL (FP32 and FP16), plus Linux arm64 Snapdragon builds targeting CPU, Adreno GPU, and Hexagon NPU.
- Android: Android arm64 (CPU) and Android arm64 Snapdragon builds supporting CPU, Adreno GPU, and Hexagon NPU.
- Windows: Windows x64 and arm64 CPU, Windows arm64 OpenCL Adreno, CUDA 12 (CUDA 12.4 DLLs on x64), CUDA 13 (CUDA 13.5 DLLs on x64 and arm64), Vulkan (x64 and arm64), OpenVINO (x64), SYCL (x64), and ROCm 10.0 (x64).
- openEuler: Specific targets for openEuler x86 and aarch64 on 310p and 910b (ACL Graph), alongside a top-level openEuler status marked DISABLED.
Developer Workflow and Inference Deployment Impact
For maintainers embedding llama.cpp inside desktop runners, developer tools, or edge-inference services, b12556 underscores why loader-level integer hygiene is critical in open-weight ecosystems. Workflows that automatically fetch community checkpoints or user-supplied models implicitly treat model headers as untrusted input. Patching #32601 reduces the risk that malformed HRM parameters crash host applications or corrupt process memory on startup across heterogeneous CPU, GPU, and NPU deployments.
From a DevOps perspective, the breadth of precompiled artifacts—spanning CUDA 12.4, CUDA 12.8, CUDA 13.5, ROCm 10.0, Vulkan, OpenVINO, SYCL, zDNN, ACL Graph, and Snapdragon NPU targets—helps teams standardize testing without maintaining local cross-compilation toolchains for every accelerator. Furthermore, combining signed commits (key B5690EEEBB952194) with GitHub attestations and bundled UI Assets gives downstream integrators a verifiable provenance trail when automating binary ingestion into staging pipelines.
Pre-Release Status, Hardware Exclusions, and Verification Caveats
Operators evaluating b12556 should note several practical caveats directly visible in the release metadata. First, the tag is marked as a Pre-release rather than a finalized stable release, making staged verification advisable before replacing production runtimes. Second, hardware coverage is not uniform: macOS Apple Silicon (arm64, KleidiAI enabled) is disabled in this build, and the openEuler matrix displays a disabled parent marker alongside its four 310p and 910b targets.
Finally, the release notes contain no throughput benchmarks, memory-footprint measurements, or formal CVE or CVSS severity ratings for #32601, so no independent performance claims can be inferred from this tag alone. The release page excerpt also omits the repository's open-source license text and vendor SDK terms; integrators redistributing binaries tied to CUDA, ROCm, OpenVINO, SYCL, or Snapdragon drivers should verify upstream repository licensing and proprietary runtime agreements independently.
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 2026-11-20.



