📁 Frameworks

The vLLM team introduced vLLM-Omni, a system designed for any-to-any multimodal models handling text, images, video, and audio. The architecture includes stage-based graph decomposition, per-stage batching, and flexible GPU allocation, achieving up to 91.4% JCT reduction tested with Qwen-Image-2512.

The first beta release of Krita 6.0 is now available, a featureful digital painting program, re-based against the Qt6 toolkit. Krita 5.3 Beta is also being released for those sticking to Qt5. The update introduces improvements in color management and Wayland support.

Intel ISPC 1.30 is now available, featuring AMX (Advanced Matrix Extensions) support added to the standard library. ISPC is a variant of the C programming language designed to target Intel CPUs and GPUs.

A developer created AnyTTS, a system that allows using any text-to-speech (TTS) engine with various AI chat interfaces, including ChatGPT and local LLM models. The integration happens via the clipboard, simplifying TTS usage across platforms. Currently, it only supports Windows, but the code is open for adaptations.

A novel reversible deep learning model employs a conditional invertible neural network to link molecular structures and 13C NMR spectra. The network, built upon i-RevNet bijective blocks, enables spectrum prediction from structure and, conversely, the generation of structure candidates from the spectrum, addressing the one-to-many nature of spectrum-to-structure inference.

A new study explores the effectiveness of the Task-Method-Knowledge (TMK) framework to enhance reasoning and planning capabilities of Large Language Models (LLMs). Results show that TMK-structured prompting can significantly increase accuracy on complex tasks, bridging the gap between semantic approximation and symbolic manipulation.

A developer has created Codag, an open-source VSCode extension that visualizes LLM workflows directly within the development environment. It supports several frameworks such as OpenAI, Anthropic, Gemini, LangChain, LangGraph, and CrewAI, along with various programming languages.

A user replaced Claude-Code's backend with NVIDIA NIM models, leveraging a free API for LLM inference. The modification includes using Telegram as an interface and preserves reasoning tokens between tool calls, enhancing performance with models like GLM 4.7 and Kimi-K2.5. The code is modular, facilitating the integration of other providers and messaging apps.

Microsoft has announced LiteBox, a sandboxing operating system developed in Rust. Designed for security, LiteBox leverages Linux Virtualization Based Security (LVBS) to isolate the guest kernel through hardware virtualization, offering a protected environment for application execution.

The Mesa project has decided to disable the use of Link-Time Optimization (LTO) during compilation due to bugs that are difficult to identify and fix. LTO, while offering performance benefits, introduces complexities in binary debugging.

Roblox's highly anticipated 4D creation feature has officially arrived in open beta. This new feature promises to open new frontiers for developers of interactive experiences on the platform.

A pull request on llama.cpp introduces a fix for the `key_gdiff` vectorized calculation in the Qwen3Next model. The change, initially reported on Reddit, aims to improve the model's accuracy and efficiency within the llama.cpp project.

A recent thread on Reddit, within the LocalLLaMA community, has sparked a heated debate about the criticisms of Ollama, a framework for local execution of large language models (LLMs). The discussion focuses on alleged shortcomings and areas for improvement in the system.

HetCCL is a library that aims to make Nvidia and AMD AI accelerators work together within the same cluster, leveraging RDMA. This vendor-agnostic approach could simplify heterogeneous AI data centers, removing obstacles to interoperability.

A new study introduces STEMVerse, a diagnostic framework to analyze the science, technology, engineering, and mathematics (STEM) reasoning capabilities of large language models (LLMs). STEMVerse aims to overcome the limitations of current benchmarks, offering a more granular assessment and a better understanding of the gaps in the models.

A novel approach, called UNSO (Unified Newton-Schulz Orthogonalization), aims to address efficiency and stability issues in the Newton-Schulz iteration, used in optimizers like Muon and on the Stiefel manifold. The method consolidates the iterative structure, avoiding polynomial expansions and optimizing coefficients for stable convergence.

Xcode 26.3 introduces agentic coding capabilities, leveraging Anthropic's Claude Agent and OpenAI's Codex. The integration aims to enhance developer efficiency by automating complex programming tasks.

Effective context management is crucial for AI agents operating on complex, long-running tasks, in order to prevent the loss of relevant information and manage the memory constraints of large language models (LLMs). LangChain's Deep Agents SDK implements context compression techniques, including offloading large tool results and inputs to the filesystem, and summarizing the message history. Targeted evaluations validate context management mechanisms.

Apple’s Xcode IDE now supports the Claude Agent SDK. This integration may simplify the development of applications leveraging Claude's capabilities.

Apple has announced Xcode 26.3, a new version of its IDE that supports agentic coding tools like Codex and Claude Agent. The integration is enabled via Model Context Protocol (MCP), allowing AI agents to interact with external tools and structured resources, including models running locally.