Skip to content
AIBRAINWORX / OpenSourceCoder.comLOCAL-FIRST. BUILT IN THE OPEN.
Inference & Quantization / Developer & Agent Tools

llama.cpp b12528 Fixes v1/realtime Sleep State Across 40-Asset Multi-Backend Build

Pre-release build b12528 of ggml-org/llama.cpp patches sleep-state handling for the v1/realtime server endpoint while publishing 40 verified cross-platform assets across CUDA, Vulkan, ROCm, SYCL, OpenVINO, and Snapdragon targets.

Listen to this articleBrowser Voice
Ready
Official company logo or product image for llama-cpp-b12528-fixes-v1-realtime-sleep-state
Official image or organization logo from github.com; identification only, no endorsement.
Event / context: 2026-11-18. Source check: October 28, 2026.

The ggml-org/llama.cpp project has published pre-release build b12528, delivering a targeted server-side fix for the v1/realtime endpoint alongside a 40-asset cross-platform binary matrix. Published on November 18, 2026 via automated GitHub Actions workflows, the release centers on pull request #32567, which addresses sleep-state behavior inside the inference server's real-time interface.

Core Announcement and Server Fix Scope

Release b12528 corresponds to commit b044890 and carries a publication timestamp of November 18, 2026 at 22:48:02 UTC (14:48:02 PST in America/Los_Angeles), with the GitHub release entry listing automated publication by github-actions at 23:38 UTC on the same date. The sole code change highlighted in the release notes is server: fix sleep state for v1/realtime (#32567), pointing directly to lifecycle and idle-state handling in llama.cpp's HTTP and streaming server layer.

Within local and self-hosted LLM stacks, the v1/realtime endpoint supports low-latency, persistent, or session-oriented inference workflows rather than single-shot request-response completions. When a server enters or exits an idle or sleep state improperly during real-time sessions, active listeners, background polling loops, or wake-up transitions can stall or drop connections. While the release summary does not reproduce the underlying diff or root-cause analysis from pull request #32567, isolating the patch to server and v1/realtime indicates a focused reliability update rather than a model-format change or new quantization scheme.

Build Matrix, Backends, and Verified Release Specs

Beyond the server patch, b12528 demonstrates the breadth of hardware acceleration targets maintained in the official llama.cpp build pipeline. The release page lists 40 downloadable assets plus UI assets, backed by provenance attestations at GitHub Attestations 69408923 and a verified GitHub commit signature using GPG key ID B5690EEEBB952194. Official project links on the release tag also reference https://llama.app.

On Apple platforms, the matrix lists macOS Apple Silicon (arm64), macOS Intel (x64), and an iOS XCFramework package. Linux builds span Ubuntu x64 and arm64 CPU targets alongside enterprise mainframe and accelerator targets, including Ubuntu 24.04 s390x (CPU), Ubuntu 26.04 s390x (CPU), and Ubuntu 26.04 s390x with zDNN acceleration. Linux GPU and NPU packages cover Vulkan for x64 and arm64, CUDA 12 (bundled with CUDA 12.8 libraries on x64), CUDA 13 (bundled with CUDA 13.5 libraries on both x64 and arm64), ROCm 10.0 on x64, OpenVINO on x64, SYCL in both FP32 and FP16 variants on x64, and Linux arm64 builds targeting Qualcomm Snapdragon CPU, Adreno GPU, and Hexagon NPU hardware.

Mobile and Windows coverage mirrors that heterogeneous compute approach. Android distributions provide arm64 CPU builds as well as Snapdragon-specific builds utilizing CPU, Adreno GPU, and Hexagon NPU execution paths. Windows targets include x64 and arm64 CPU builds, Windows arm64 OpenCL Adreno, CUDA 12 on x64 using CUDA 12.4 DLLs, CUDA 13 on both x64 and arm64 using CUDA 13.5 DLLs, Vulkan on x64 and arm64, plus Windows x64 packages for OpenVINO, SYCL, and ROCm 10.0.

Developer Workflow and Real-Time Deployment Impact

For infrastructure engineers running local inference servers across heterogeneous fleets, the combination of a v1/realtime sleep-state fix and synchronized multi-backend binaries reduces operational friction. Teams building voice interfaces, interactive coding assistants, or persistent agentic sessions on top of llama.cpp rely on predictable server state transitions when traffic fluctuates between bursts and idle windows. Shipping the fix simultaneously across Vulkan, CUDA 12, CUDA 13, ROCm 10.0, SYCL, OpenVINO, and Snapdragon NPU builds allows developers to test consistent server behavior across workstations, edge devices, and GPU nodes without compiling from source.

The cryptographic verification metadata included with b12528 also supports reproducible supply-chain practices for open-source maintainers and enterprise evaluators. Automated build provenance via GitHub Attestations and verified GPG signing lets downstream packagers confirm that prebuilt binaries and UI assets originated from commit b044890 before integrating them into container images or internal CI pipelines.

Pre-Release Caveats, Disabled Targets, and Verification Boundaries

Several operational caveats apply before deploying b12528 in production environments. First, GitHub explicitly classifies b12528 as a pre-release rather than a stable milestone, and the tag summary provides no benchmark figures, memory footprints, or independent latency measurements for the v1/realtime fix. Engineering teams should inspect pull request #32567 directly and validate idle-to-active wake behavior against their specific model weights and concurrency profiles.

Second, two build groups in the release matrix are explicitly marked as disabled on the tag page: the macOS Apple Silicon (arm64) build with KleidiAI enabled, and the openEuler build group covering x86 and aarch64 targets for 310p and 910b (ACL Graph). Environments relying on Arm KleidiAI kernels or Ascend-focused openEuler packages will need to track subsequent build tags or compile locally if supported. Operators should also note the library-version split across CUDA 12 artifacts, where Ubuntu x64 packages list CUDA 12.8 libraries while Windows x64 packages list CUDA 12.4 DLLs. Finally, the release tag view itself does not display repository licensing terms or hardware-driver prerequisites, so teams must verify the upstream llama.cpp license, vendor runtime agreements, and target driver compatibility prior to redistribution or production rollout.

EXPLORE THE SOURCE

GitHub repositories

Automated Vertex AI (gemini-4-hydrogen) primary-source report; reviewed 2026-11-19.

Follow your curiosity.

Pass the signal.