
The open-source inference project llama.cpp has published pre-release b12530 under commit b001d7b, introducing targeted updates to its ggml WebAssembly backend alongside a full cross-platform binary build. Tagged by github-actions on November 19, 2026 at 01:10 UTC and published on GitHub at 01:13:35 UTC—November 18, 2026 in Pacific Standard Time—the llama.cpp b12530 release centers on pull request #19590, titled "ggml : WASM Relaxed SIMD Enhancement" and co-authored by Xuan Son Nguyen. While b12530 ships without standalone benchmark numbers, it fixes existing WebAssembly Relaxed SIMD kernels and adds support for q3_K and q6_K quantized formats.
Operational scope and WebAssembly kernel update
As a rolling pre-release snapshot of ggml-org/llama.cpp, b12530 focuses its single documented source change inside the ggml tensor engine rather than altering high-level application APIs or model formats. According to the commit log accompanying pull request #19590, the patch series adds relaxed SIMD support, corrects existing relaxed SIMD kernel implementations, enables q3_K and q6_K quantization coverage in WebAssembly, and performs codebase cleanup across indentation, comments, and trailing whitespace. In quantized local inference workflows, q3_K and q6_K represent distinct K-quant block bit-widths used to compress large language model weights while balancing memory footprint and precision. Extending corrected Relaxed SIMD kernels to those formats allows WebAssembly builds of ggml to process q3_K and q6_K tensors using modern vector instructions instead of relying solely on older baseline paths. Beyond the core C and C++ kernel change, the release packages 40 UI Assets and publishes automated build attestations for downstream verification.
Engineering components and multi-backend build matrix
Commit b001d7b carries a verified GitHub signature using GPG key ID B5690EEEBB952194 and links to build attestation 69435594 as well as the companion web address llama.app. Alongside the WebAssembly update, the automated release pipeline compiles binaries spanning desktop, mobile, workstation, and server accelerators. On Apple hardware, b12530 provides macOS Apple Silicon (arm64), macOS Intel (x64), and iOS XCFramework packages, while the macOS Apple Silicon (arm64, KleidiAI enabled) variant is listed as DISABLED.
Linux coverage spans Ubuntu x64 (CPU), Ubuntu arm64 (CPU), Ubuntu 24.04 s390x (CPU), Ubuntu 26.04 s390x (CPU), and Ubuntu 26.04 s390x (zDNN), paired with GPU and accelerator builds for Ubuntu x64 and arm64 (Vulkan), Ubuntu x64 (CUDA 12) with CUDA 12.8 libraries, Ubuntu x64 and arm64 (CUDA 13) with CUDA 13.5 libraries, Ubuntu x64 (ROCm 10.0), Ubuntu x64 (OpenVINO), Ubuntu x64 (SYCL FP32 and SYCL FP16), and Linux arm64 (Snapdragon: CPU, Adreno GPU, Hexagon NPU). Mobile and Windows targets include Android arm64 (CPU), Android arm64 (Snapdragon: CPU, Adreno GPU, Hexagon NPU), Windows x64 and arm64 (CPU), Windows arm64 (OpenCL Adreno), Windows x64 (CUDA 12) with CUDA 12.4 DLLs, Windows x64 and arm64 (CUDA 13) with CUDA 13.5 DLLs, Windows x64 and arm64 (Vulkan), Windows x64 (OpenVINO), Windows x64 (SYCL), and Windows x64 (ROCm 10.0). Four openEuler entries targeting x86 and aarch64 on 310p and 910b (ACL Graph) hardware also appear under a DISABLED section marker.
Workflow impact for browser and edge developers
For developers deploying llama.cpp inside browser tabs, Progressive Web Apps, or sandboxed WASI runtimes, fixing WASM Relaxed SIMD kernels directly addresses execution reliability on modern JavaScript and WebAssembly engines. Adding q3_K and q6_K support is particularly relevant for client-side deployments where bandwidth and browser heap limits matter: teams that distribute 3-bit or 6-bit K-quantized GGUF weights can target a unified Relaxed SIMD kernel path rather than maintaining separate quantization rules for WebAssembly versus native binaries. Meanwhile, the breadth of the b12530 matrix—stretching from CUDA 12.4 and 13.5 toolchains to ROCm 10.0, Vulkan, SYCL, OpenVINO, IBM zDNN on s390x, and Qualcomm Snapdragon Hexagon NPUs—helps cross-platform engineering teams test a single pinned commit across cloud GPUs, developer laptops, and edge clients. The inclusion of verified GPG signatures and GitHub build attestations also gives build engineers a cryptographic trail to verify that prebuilt artifacts correspond directly to commit b001d7b.
Limitations, verification gaps, and deployment caveats
Despite its wide distribution footprint, b12530 remains explicitly flagged as a GitHub pre-release, meaning production adopters should treat it as an upstream integration snapshot rather than a long-term support milestone. Crucially, the release notes publish no tokens-per-second throughput measurements, latency figures, memory benchmarks, or accuracy comparisons for the revised WASM Relaxed SIMD kernels or the new q3_K and q6_K paths; developers must run their own hardware and browser profiling before assuming speed gains. Operational portability also depends on runtime support for WebAssembly Relaxed SIMD, which can expose architecture-specific rounding or instruction differences across host CPUs compared with strict deterministic SIMD modes. In addition, the DISABLED markers on the KleidiAI-enabled Apple Silicon build and the openEuler section leave teams on those specialized targets without active binaries in this tag. Finally, because the release page itself does not reproduce full repository licensing text or independent security evaluations, organizations should inspect the upstream license terms and validate attestation 69435594 prior to production 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 2026-11-19.



