Models & Labs
Llama.cpp b10093 Release Fixes DeepSeek4 Template
llama.cpp Releaseshigh confidence
Why it matters
- →Fixes to DeepSeek4 template improve reliability for developers.
- →Enhanced support for various platforms ensures broader usability.
- →Integration with technologies like Vulkan and CUDA boosts performance.
Llama.cpp has released version b10093, focusing on fixing the DeepSeek4 crafted template to ensure it follows reference behavior. The update also supports the DeepSeekv4 flag and hooks the DS3.2 parser for DS4, enhancing the tool's functionality. This release is available across multiple platforms, including macOS, Linux, and Windows, with specific support for technologies like Vulkan, ROCm, and CUDA. The update aims to improve performance and stability for developers using these environments.
Read originalMore from llama.cpp Releases
Models & Labsmodels
llama.cpp b10094 Release Enhances Speculative Type Inference
The latest llama.cpp release, version b10094, refines how speculative type inference is handled for draft repositories. By automatically identifying sidecars and prioritizing them according to a set hierarchy, users can now resolve drafts more seamlessly without additional flags. This update streamlines the workflow for developers dealing with intricate model setups, ensuring the correct draft is chosen more effectively. The release continues to support multiple platforms, including macOS, Linux, Windows, and openEuler, making it a flexible tool for developers working in different environments.
llama.cpp Releases
Models & Labsmodels
Llama.cpp b10098 Release Enhances Hexagon Ops
The b10098 release of llama.cpp brings a series of enhancements to Hexagon activation operations, focusing on performance optimization with new geglu microkernels and support for non-contiguous source and strided DMA. These updates aim to refine the efficiency of GLU per-thread functions and reorganize unary operations, potentially boosting computational performance on various platforms. While the release doesn't introduce groundbreaking features, it strengthens llama.cpp's capabilities for developers working with diverse hardware setups. This update reflects llama.cpp's ongoing efforts to optimize AI inference across multiple systems, making it a more reliable tool for developers.
llama.cpp Releases
Models & Labsmodels
Llama.cpp b10099 Release Enhances CUDA Quantization
The latest b10099 release of llama.cpp brings notable improvements to CUDA quantization, specifically enhancing NVFP4 W4A4 activation quantization. This update introduces optimizations like 32-byte loads and the use of nvfp4x4 intrinsic, which streamline performance and efficiency. By fusing per-channel amax and quantization kernels, the release aims to boost computational efficiency. While the update doesn't introduce new models, it refines the existing infrastructure, making it more robust for developers working with NVIDIA GPUs. This release is a step forward in optimizing AI model performance on CUDA platforms.
llama.cpp Releases
More in Models & Labs
Models & Labsmodels
Claude Code v2.1.219 Update Released
Claude Code's latest update, v2.1.219, introduces the Claude Opus 5 model as the new default, offering a 1M context and fast mode pricing. This release enhances sandbox security with a strict allowlist setting and improves error handling with structured failure categories. Notably, nested subagent forwarding is now supported up to a depth of three, expanding the tool's flexibility. These changes make Claude Code more robust and versatile, particularly for developers managing complex workflows.
Claude Code Releases
© TechCrunch AIModels & Labscoding
OpenAI Launches AI Keypad for Coders
OpenAI has ventured into hardware with the launch of Micro, a keypad designed to enhance the ChatGPT experience. Developed with Work Louder, this device targets tech enthusiasts and coders, offering customizable keys for specific tasks within ChatGPT and Codex. While the keypad includes features like voice dictation and customizable commands, its $230 price tag and niche appeal have drawn mixed reviews. For those deeply integrated with ChatGPT, Micro could streamline workflows, but its utility may not justify the learning curve for casual users.
TechCrunch AI
© Lev SelectorModels & Labsmodels
Alibaba Previews Qwen 3.8 Max Model
Alibaba has released a preview of its new Qwen 3.8 Max model.
Lev Selector