

The llama.cpp b12490 pre-release introduces a targeted runtime fix for qwen4exp, resolving a computation graph reallocation issue tied to n-gram table scaling.
What Changed in b12490
Tagged from commit ad9d9e8 and published via GitHub Actions on November 17, 2026 (Pacific Time; November 18, 2026 at 01:17 UTC), the release centers on pull request #32483. Instead of triggering graph reallocation during n-gram table scaling, the update reads the scale once after loading and applies it directly through ggml_scale, alongside minor comment cleanup.
Build Matrix and Artifact Notes
Release notes list build attestations and binaries spanning macOS (Apple Silicon and Intel), iOS XCFramework, Linux (Ubuntu x64, arm64, and s390x across CPU, Vulkan, CUDA 12.8/13.5, ROCm 10.0, OpenVINO, SYCL, zDNN, and Snapdragon targets), Android arm64, and Windows (x64 and arm64). According to the release manifest, the macOS Apple Silicon KleidiAI-enabled target and openEuler builds are marked disabled in this pre-release. No independent performance benchmarks are included in the release notes.
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.



