Models & Labs
llama.cpp b11025 adds CUDA 13 and ROCm 10.0 support
llama.cpp Releaseshigh confidence
Why it matters
- →CUDA 13 and ROCm 10.0 pre-builts eliminate compilation barriers for users with latest-gen NVIDIA and AMD GPUs.
- →OpenEuler Ascend 910b support expands local inference options beyond Western hardware ecosystems.
- →KleidiAI disabled by default suggests stability prioritization over raw performance gains on Apple Silicon.
llama.cpp has released version b11025, introducing pre-built binaries for CUDA 13 and ROCm 10.0 across Linux and Windows platforms. The update also adds support for openEuler with Huawei Ascend 910b ACL Graph acceleration, broadening compatibility with Chinese AI hardware ecosystems. Notably, KleidiAI optimizations on macOS Apple Silicon are disabled by default in this build, while Vulkan and OpenVINO support remain consistent across major operating systems. This release addresses the growing need for support of newer GPU architectures without requiring users to manually compile the library.
Read originalMore from llama.cpp Releases
Coding Toolscoding
llama.cpp b11017 adds CUDA 13 and ROCm 10 builds
This release quietly extends llama.cpp's hardware support to the latest NVIDIA and AMD stacks. By shipping native binaries for CUDA 13 and ROCm 10, it ensures compatibility with newer GPU architectures without requiring users to compile from source. The inclusion of both CUDA versions side-by-side is a pragmatic move for enterprise environments managing mixed infrastructure. While no new model formats are introduced, this update keeps the runtime relevant as hardware vendors push their latest drivers.
llama.cpp Releases
Coding Toolscoding
llama.cpp b11018 adds CUDA 13 and ROCm 10 builds
This release quietly cements llama.cpp as the universal inference runtime by finally supporting NVIDIA's latest CUDA 13 stack alongside AMD's ROCm 10. For the first time, users on cutting-edge hardware can run local models without being forced into legacy driver versions or waiting for vendor-specific optimizations. The inclusion of both CUDA 12 and 13 binaries side-by-side removes a major friction point for developers managing mixed environments. While no new model architectures are added, this infrastructure update ensures compatibility with the fastest consumer and data center GPUs hitting the market right now.
llama.cpp Releases
Coding Toolscoding
llama.cpp b11019 fixes embedded GGUF loading bugs
This release patches a subtle but critical bug where embedding a GGUF file inside another container caused tensor data corruption due to misaligned memory mapping. The fix ensures the data section is aligned relative to the GGUF header rather than the outer file offset, preventing silent errors that could ruin inference results. It also adds support for loading LoRA adapters via FILE pointers and disables mmap with a warning when alignment issues are detected. This stabilizes llama.cpp for complex packaging scenarios where models are bundled within larger archives or custom formats.
llama.cpp Releases
More in Models & Labs
© TechCrunch AIInvestment · $22.25M
Models & Labsmodels
PrismML compresses LLMs to run locally with minimal loss
PrismML is proving that extreme model compression doesn't have to mean dumb models. Their Bonsai 2 27B model shrinks Alibaba's Qwen3.8 down to just 5.9 GB using ternary weights, hitting 98% of the original benchmark scores. This isn't just a technical curiosity; it means high-performance reasoning can finally run on consumer hardware without cloud dependency. With $22.25M in seed funding and backing from Khosla Ventures, they are positioning themselves as the bridge between massive lab models and private, local inference.
TechCrunch AI
© TechCrunch AIModels & Labsmodels
Huawei advances Ascend 960DT launch to Q1 2027
Huawei is accelerating its next-generation Ascend 960DT chip to Q1 2027, cutting the original timeline by six months. This move signals a push to scale its Peerium Computing Architecture, which aims to link hundreds of thousands of accelerators into a single massive system. While the chip release is earlier than expected, there are conflicting reports about the scaling capacity of their Atlas SuperCluster, with some analysts noting a reduction in claimed node counts. The shift highlights Huawei's determination to build domestic AI infrastructure despite U.S. export restrictions. It marks a tangible step in China's effort to reduce reliance on Nvidia for large-scale training and inference workloads.
TechCrunch AI
© The AI Daily BriefModels & Labsmodels
TypeSafe Launches Jev Judgment Model
TypeSafe introduces Jev, a model that outputs calibrated probabilities instead of text, claiming 20-200x speed improvements over traditional LLMs for decision tasks.
The AI Daily Brief