Models & Labs
llama.cpp b11154 improves testing and adds CUDA 13
llama.cpp Releaseshigh confidence
Why it matters
- →CUDA 13 support allows developers to test against NVIDIA's newest API version early.
- →ROCm 10.0 binaries keep AMD GPU users aligned with recent driver releases.
- →Improved test output helps maintainers catch regression issues faster across platforms.
llama.cpp has released build b11154, focusing on infrastructure updates and platform support. The release adds CUDA 13 binaries for Linux and Windows, alongside ROCm 10.0 support for AMD GPUs. Internal testing logic was refactored to provide clearer pass/fail reporting during multi-model validation. KleidiAI optimizations for Apple Silicon are temporarily disabled in this build. The update ensures compatibility with the latest driver ecosystems for local inference.
Read originalMore from llama.cpp Releases
Coding Toolscoding
llama.cpp b11140 optimizes sparse attention on CUDA
This release targets a specific bottleneck in long-context inference by optimizing the sparse flash attention prefill step for NVIDIA GPUs. By templating kernels to unroll loops at compile time, batched sparse operations drop from 586 microseconds to 244 microseconds on 49k context windows. This isn't just a generic speed bump; it makes handling very long documents significantly more efficient for users relying on sparse attention mechanisms. The change is already baked into the standard CUDA builds, requiring no special flags.
llama.cpp Releases
Coding Toolscoding
llama.cpp b11146 adds CUDA 13 and Snapdragon support
The latest llama.cpp build brings immediate relevance to users on bleeding-edge NVIDIA hardware with native CUDA 13.4 support across Linux and Windows, closing the gap for those testing next-gen GPU architectures. More notably, it finally addresses the mobile inference landscape by including a dedicated build for Linux arm64 Snapdragon devices, covering CPU, Adreno GPU, and Hexagon NPU paths. This moves local AI beyond just desktop GPUs into the realm of high-performance edge computing on Qualcomm silicon. While Apple Silicon builds have KleidiAI disabled in this specific release, the expansion to ARM-based mobile NPUs marks a significant shift in where llama.cpp can run efficiently.
llama.cpp Releases
Coding Toolscoding
llama.cpp b11147 adds ROCm 10 and Snapdragon support
This release quietly cements llama.cpp as the universal inference runtime by finally bringing first-class ROCm 10.0 support to both Linux and Windows. AMD GPU users no longer need workarounds, effectively closing a long-standing parity gap with NVIDIA's CUDA ecosystem. Equally notable is the new Linux arm64 build targeting Snapdragon chips, which unlocks local LLM execution on high-performance mobile hardware via CPU, Adreno GPU, and Hexagon NPU acceleration. While KleidiAI on Apple Silicon has been disabled in this specific binary set, the broader platform expansion signals a shift toward heterogeneous computing that extends well beyond traditional desktop GPUs.
llama.cpp Releases
More in Models & Labs
Models & Labscoding
vLLM v0.30.1rc0 adds AMD MI355 NVFP4 support
vLLM is quietly closing the hardware gap for AMD users with this release candidate. By adding dense NVFP4 and MoRI kernel mirrors for the new MI355 GPU, they are enabling high-efficiency inference on hardware that previously lacked first-class support. This isn't just a driver update; it's a critical infrastructure patch that allows enterprises to deploy advanced quantization formats on AMD silicon without waiting for upstream integration. The inclusion of OpenAI Codex in the commit history suggests automated testing is helping maintain this parity, making AMD a more viable option for cost-sensitive inference workloads.
vLLM Releases
© The Verge AIModels & Labsother
Meta launches camera-free Ray-Ban Meta Audio Glasses
Meta is betting that privacy concerns are the biggest barrier to smart glasses adoption, so it launched a version with zero cameras. These audio-only Ray-Bans strip out the recording hardware to drop the weight to 43 grams and the price to $349, targeting users who want Meta AI assistance without the 'pervert' stigma. By removing the camera, Meta could rearchitect the frame for better design flexibility while keeping battery life up to 12 hours. This move signals a strategic pivot toward mass-market comfort over surveillance capabilities, acknowledging that audio-first interaction is the most viable entry point for mainstream wearables.
The Verge AI
© The Verge AIModels & Labsother
Meta unveils VR glasses and camera-free AI wearables
Meta’s hardware strategy is pivoting hard toward lightweight, socially acceptable AI wearables. The headline grabber is the Meta VR Glasses: a 100g form factor that offloads compute to a pocket puck, effectively solving the bulk problem that has plagued mixed reality for years. Simultaneously, they launched camera-free Ray-Ban Meta Audio Glasses to neutralize the 'pervert glasses' backlash while keeping AI features alive. The Muse agent is also getting its own dedicated hardware, the Muse Charm, signaling a shift from app-based agents to always-on physical interfaces.
The Verge AI