Qwen-Image-Flash, Anthropic Oceanus, Apple Poke AI agent

Researchers distilled Qwen-Image-2.0 into a lightweight model, Qwen-Image-Flash, using a few-step training recipe. They show that data composition, teacher guidance, and task mixing crucially boost student performance, enabling rapid visual generation and instruction‑guided editing with far fewer steps.

A new model identifier “claude‑oceanus‑v1‑p” surfaced in Anthropic’s console and was given to internal red‑team testers. The Medium report notes the leak and suggests Anthropic is gearing up for a public launch of the next Mythos‑class model within weeks.

A new arXiv study directly compares diffusion probabilistic models and flow‑matching (rectified flow) using the same U‑Net backbone on MNIST. It finds that flow‑matching achieves comparable image fidelity with far fewer function evaluations, making it far more efficient for low‑resource hardware. The analysis highlights a smoother transport path and an efficiency frontier at just ten steps.

Poke becomes the first third‑party AI agent approved for Apple’s Messages for Business platform, letting iMessage users chat with an autonomous assistant for tasks such as calendar management, health tracking, and smart‑home control. The startup will pay Apple a per‑user fee, creating a new revenue channel for the tech giant.

President Trump issued a National Security Presidential Memorandum that accelerates AI adoption across the Department of Defense and intelligence agencies, mandating rapid onboarding of leading AI models, building secure high‑performance computing, and creating an AI talent reserve. The memo replaces the previous administration’s AI framework.

The official NeurIPS 2026 main‑track handbook explicitly prohibits authors from embedding hidden prompts aimed at manipulating LLM‑based reviewers, labeling such prompt‑injection attacks as a breach of review policy. This warning targets reciprocal reviewing practices and follows similar concerns raised at ICML.

A new pull request to the llama.cpp repository adds support for DeepSeek V4 Flash, enabling the model to run on the C++ inference engine. Although still experimental and slow, it demonstrates the first functional integration of the DeepSeek V4 series with llama.cpp, requiring high‑VRAM GPUs.

OpenLumara is an open‑source, GPL‑2 AI agent framework written from scratch in Python, built for local LLMs (llamacpp, koboldcpp, Ollama). Its token‑efficient design uses a ~4k‑token system prompt and modular components that can be toggled, offering WebUI, CLI, Telegram, Discord, and Matrix integrations.

The EVA‑Bench Data 2.0 release adds three enterprise voice‑agent domains, Airline Customer Service, IT Service Management, and Healthcare HR Service Delivery, covering 213 realistic scenarios and 121 tooling configurations. All datasets are open‑source on Hugging Face and have been validated against leading LLMs, enabling robust benchmarking and dataset creation.