C++ news and engineering summaries

llama.cpp is an open-source C/C++ project enabling local and cloud LLM/VLM inference with minimal setup and high performance. Built on ggml, it supports extensive CPU, GPU, and accelerator backends, integer quantization, hybrid CPU-GPU execution, Docker and source installation, Hugging Face models, and OpenAI-compatible REST APIs through llama-server.

WhatsApp has successfully integrated and rolled out a significant security layer built with the Rust programming language, marking a major milestone in production-grade software engineering. This initiative, known as Kaleidoscope, replaces a legacy C++ library with 90,000 lines of Rust code to monitor and validate media files across billions of devices. By leveraging Rust's memory-safe properties, WhatsApp aims to proactively neutralize vulnerabilities like the 2015 Stagefright exploit, which targeted media parsing libraries. This rollout spans Android, iOS, Windows, Mac, and web platforms, proving Rust's scalability and performance advantages in a diverse, global environment. The transition highlights a broader shift toward memory-safe languages to mitigate high-severity vulnerabilities common in C and C++ implementations. WhatsApp's strategy underscores a defense-in-depth approach, ensuring end-to-end encryption is complemented by robust protection against sophisticated malware hidden in seemingly benign file types.

Microsoft has open-sourced Comic Chat, the 1996 IRC client famous for introducing the Comic Sans font. By transforming text conversations into real-time comic panels with automated characters and expressions, it pioneered visual communication. The source code, written in C++ and MFC, is now available on GitHub for historical study and modernization on modern Windows systems.

AnyPS5 automatically ports PS5 executables to Linux and Windows without emulation or a separate runtime process. Its relinker converts executables to native target formats, while implementations of system prx libraries support dynamic linking. The project includes SDL-mapped controller input, configurable keyboard and mouse controls, and a shader recompiler that produces validated SPIR-V. Dreaming Sarah reaches stable 60 fps on a GTX 1050 Ti and i5-7500; unsupported states terminate with std::runtime_error.

GOG has officially announced its intention to bring GOG Galaxy, its gaming client and ecosystem hub, natively to Linux. Describing the operating system as the 'next major frontier,' the company is currently recruiting a senior engineer to oversee the architectural transition of its large C++ codebase. This strategic move follows a leadership shift where GOG's original co-founder regained control of the company with a vision to compete against Steam by prioritizing quality over quantity. While technical solutions like Proton have already improved Linux compatibility, a native GOG Galaxy client represents a significant step forward in providing seamless library management and community features for Linux users. The initiative aims to lower barriers for enthusiasts wanting to play retro classics and newer titles without the typical hurdles of non-native software layers, strengthening the viability of FOSS platforms for mainstream gaming.

DeepGEMM is a unified CUDA library of high-performance tensor-core kernels for LLM primitives, including FP8, FP4, and BF16 GEMMs, grouped and masked MoE operations, MQA scoring, HyperConnection, and Mega MoE. It uses DeepJIT to compile kernels at runtime, avoiding CUDA compilation during installation, and favors a small, accessible codebase over heavy CUTLASS templates. It targets NVIDIA SM90 and SM100 GPUs and claims performance matching or exceeding expert-tuned libraries, reaching up to 1550 TFLOPS on H800.

Hayder Tirmazi details four llama.cpp optimizations for prompt lookup drafting: eliminating unnecessary map copies, replacing maps with faster data structures, and using constmap for the static cache. Together, they deliver up to 42× faster drafting and 2.6× lower memory use, while preserving acceptance rates; static-cache loading also becomes up to 16× faster.

The article develops a C++ single-producer, single-consumer (SPSC) ring buffer from a basic FIFO queue to a lock-free design. Replacing a mutex with atomics, tuning memory ordering, and caching the other thread’s index raises reported throughput from 12M to 305M ops/s (25×); results depend on the benchmark setup and hardware.

Tether is an MIT-licensed Linux companion for iPhone that recreates technically feasible macOS Continuity features: iMessage and SMS, notifications, contacts, file transfer, clipboard sync, and OTP autofill. Built with secure mTLS and a clean-room C++ implementation of Bluetooth interoperability, it addresses Linux–iPhone gaps and welcomes contributions.

Rad-Therapy II is an ambitious open-source project designed to port Valve Corporation's Half-Life 2 (2004) into the Quake engine ecosystem, specifically utilizing FTEQW and the Nuclide SDK. Although the game is not currently playable from start to finish, it supports various multiplayer modes like deathmatch. The project requires original assets from Half-Life 2 and Half-Life 2: Deathmatch to function lawfully. Technically, the repository manages game logic, GLSL shaders, and plugins necessary for the engine to interpret Source engine data files. Licensed under the ISC License, the project represents a significant intersection between the legacy QuakeWorld community and modern game asset reverse-engineering.

A guide for aspiring graphics programmers emphasizing the duality of rendering: CPU-side API mastery (DirectX12, Vulkan) and GPU-side rendering techniques (path tracing, PBR). The author recommends C++ as the industry standard, highlights essential linear algebra and algorithm skills, and suggests a cautious, practical approach to integrating machine learning into development workflows.

C++: The Documentary recently premiered on YouTube, detailing the 40-year history of the language from its origins at Bell Labs to its status as a top-tier industry standard. Featuring industry pioneers like Bjarne Stroustrup, the film highlights the evolution, standardization, and massive recent growth of C++ in fields like gaming, physics, and finance.

A high-efficiency C++/CUDA inference engine for Llama models, optimized for running 70B parameters on consumer GPUs like the RTX 3090. It features 3-tier adaptive caching, NVMe direct I/O for bypassing the CPU, and layer skipping via cosine similarity. The engine achieves significant speedups through double-buffered streaming and self-speculative decoding without external dependencies.

The ISO C++ committee has completed technical work on C++26. This major release features compile-time reflection, significant memory safety enhancements (including hardened standard library operations), standardized contracts for functional safety, and the std::execution framework. Industry adoption is expected to be rapid, with work already commencing on C++29 to further advance memory safety and language security.

A prominent Carnegie Mellon professor is joining ClickHouse to lead ClickHouse Labs, a new research division. The team aims to advance database technology by bridging academic research and industrial application. Primary focuses include innovating ClickHouse and PostgreSQL architectures, optimizing performance, and exploring the integration of DBMS solutions with emerging AI and agentic technologies.

Developer Dante Leoncini successfully ported Half-Life to the Nokia N95, a 2007 Symbian-based smartphone. The game runs at 30 FPS with added Bluetooth mouse and keyboard support. This accomplishment highlights the capability of legacy mobile hardware, using a native Symbian build rather than emulation to achieve performance comparable to 1990s desktop PCs.

ClickHouse, which began as a 2009 internal tool for web analytics, has become a leading open-source analytical database. Built from scratch in C++, it emphasizes modular, well-documented code and a transparent development process. Its evolution from custom prototypes to a high-performance, globally adopted DBMS demonstrates the power of open-source collaboration and engineering discipline.

DSPi is an affordable, high-performance digital audio processor for Raspberry Pi Pico and Pico 2 boards. It functions as a USB sound card, providing tools like parametric EQ, room correction, volume leveling, loudness compensation, and headphone crossfeed via a low-latency C++ signal chain suitable for diverse audio configurations using S/PDIF, I2S, and PDM outputs.

llama.cpp v0.6.0 adds the llama_batch_ext API and llama_process() for mixed token/embedding batches and MTP/deepstack state embeddings. It supports GLM-5.3-Flash, Clef, Ling 3.0 VL, Nimble, and Qwen4Exp MTP speculative decoding, which delivers about 1.5× faster decoding on DGX Spark. The release introduces the /v1/systemone decision-model API, a Hugging Face Hub-powered Web UI and model downloads, new Metal and Vulkan attention kernels, and ggml v0.26.0.

The C++ standard library has struggled for 15 years with design flaws it cannot easily fix due to strict ABI compatibility requirements. As a result, the library has become a collection of suboptimal defaults—like legacy containers and wrappers—widely discouraged in modern codebases, forcing professional engineers to rely on third-party libraries instead of standard solutions.