windows

LlamaStation v0.9 is a Windows GUI for llama.cpp that offers a clean interface with full parameter control, multiple backends (official, TurboQuant, AtomicChat, BeeLlama), real-time VRAM monitoring, per-model profiles, voice mode, and headless mode, all without intermediate layers like Ollama.

This article from The Old New Thing explains the concept of making APIs 'inert'—do nothing in a way that avoids breaking existing apps—using examples like printing support on Xbox and retiring widget APIs.

An exploration of the origin of the ERROR_ARENA_TRASHED error code from MS-DOS, its meaning in memory management, and why it's no longer used in Win32 but persists as a legacy code.

Microsoft launches Surface Pro 12 and Surface Laptop 8 for businesses with Intel Core Ultra Series 3 processors, offering higher RAM options and 5G connectivity, but at significantly increased prices.

QuestDB engineers debug a sporadic Windows hang caused by a deadlock involving the Windows DLL Loader Lock, Rust thread-local storage destruction, JNI detach, and JVM garbage collection safepoint mechanics.

Community benchmarks of Qwen 3.6-27B Dense and MTP variants running via llama.cpp on Strix Halo Windows, showing token/s speeds for various tasks.

Hunk version 0.12.0 is live with new installation options (Homebrew, Nix), lazygit support, scroll performance improvements, and Windows support.

CVE-2026-40369 describes a vulnerability in Windows kernel's NtQuerySystemInformation function that allows arbitrary kernel address increment, enabling privilege escalation from unprivileged processes including Chrome sandbox. The exploit is deterministic on Windows 11 24H2-25H2.

A user shares a fix for performance bottlenecks when running AI models on AMD GPUs in Windows 11 by disabling memory compression via the command 'Disable-mmagent -mc'.

An open-source terminal tool called witr, developed by an Indian independent developer, explains why a process is running, traces its origin from kernel to systemd, and includes an interactive TUI dashboard. It gained 15,400+ GitHub stars in less than five months.

OpenAI is improving agent support on Windows by implementing a custom sandbox for Codex, addressing OS-level isolation challenges to ensure safe and efficient operation.

A debugging story about a Windows program hanging when the user changes keyboard layouts due to a background thread that created a window but wasn't pumping messages. The fix is to either pump messages or destroy the window.

OpenAI engineers built a custom sandbox for Codex on Windows to enable safe, restricted command execution, balancing effectiveness and security without relying on native Windows isolation features.

The user shares a report on successfully running the Qwen3.6 35B A3B model on Windows using Hermes Agent and an 8GB VRAM GPU.

Explains the historical reason why 32-bit Windows client editions artificially limit RAM to 4 GB: driver compatibility issues with Physical Address Extensions (PAE) and Data Execution Prevention (DEP), as opposed to any nefarious motive.

Raymond Chen describes a technique using a helper process to precisely control which handles are inherited by a new process, avoiding accidental inheritance from other components in the same process.

The Hermes AI agent tool now supports native Windows installation without requiring WSL, allowing Windows users to run CLI tools and bots directly. This article reviews the Early Beta experience, outlines the installation process, and compares it with the WSL2 version.

Asteroid is a computer use agent builder that works across Browser, Linux, and Windows, enabling users to create automated agents.

OpenLess is an open-source voice input tool that supports macOS and Windows, capable of converting speech to text and automatically polishing it, especially suitable for writing AI prompts.

The article examines the pitfalls of altering API behavior depending on the linked SDK version, using Windows' CoInitializeSecurity as a case study. It discusses issues with DLL version mismatches and tail call optimization that complicate this approach.