Xiaomi Releases 1T-Parameter MoE Model MiMo-V2.5-Pro

Xiaomi has released MiMo-V2.5-Pro, an open-source Mixture-of-Experts (MoE) language model. It has a total of 1.02T parameters with 42B active, and adopts a hybrid attention architecture along with 3-layer Multi-Token Prediction (MTP). The maximum context length reaches 1 million tokens.

Xiaomi positions it as its "most powerful model," saying it was designed with the most demanding use cases in mind, such as agentic tasks, complex software engineering, and long-running task execution. The arrival of a 1T-class open-source MoE model from a Chinese vendor can be seen as a sign of the intensifying open-source AI competition.

Apple Agrees to $250M Settlement Over Misleading Ads About Siri's AI Features

Apple has submitted a settlement proposal to a US court in a class-action lawsuit over misleading advertising of its voice assistant Siri's AI features, agreeing to pay $250 million (about 39 billion yen).

The issue stems from development delays in Apple Intelligence, the AI platform Apple announced in 2024. The company heavily promoted advanced AI capabilities for Siri, but their actual delivery lagged significantly, allegedly betraying consumer expectations. If the settlement is approved, eligible US consumers are expected to receive a base payout of $25 per device.

The case shows that consumers are keeping a closer eye on overblown promotion of AI features—not only from AI companies but also from platform companies.

Google Quietly Removes On-Device AI Privacy Guarantee

It has emerged that Google quietly removed language guaranteeing on-device AI privacy in the process of embedding AI models into its devices.

Google has long emphasized that completing AI processing on the device keeps user data from being sent to the cloud, but in reality the company appears to be shifting toward a setup that does send data to the cloud in order to improve AI model capabilities. Changing privacy promises after the fact has stirred discussion on Hacker News as something that undermines user trust.

"Think Step by Step" Backfires on Local LLMs — Observing the Limits of CoT

An interesting experimental result was reported on Qiita. The author observed three scenarios in which Chain-of-Thought (CoT) prompts like "Let's think step by step" actually hurt accuracy on smaller models such as Qwen2.5-14B running in a local environment (8GB VRAM).

Since a 2022 paper, CoT has been widely accepted as a "magic spell" that improves LLM accuracy, but this report suggests that CoT can be counterproductive when the model is small. In particular, models with limited reasoning ability may actually get confused by step-by-step instructions.

As a practical demonstration of the lesson that "no universal technique works for every model" in prompt engineering, this is a useful reference for anyone working with local LLMs.

A Tool for Running AI Coding Agents in Isolated Docker Containers

An open-source tool for running AI coding agents inside isolated Docker containers has been released on GitHub.

When AI agents autonomously generate and execute code, preventing unintended effects on the host system is essential from a safety perspective. This tool runs each agent session in an independent container, sandboxing filesystem access and command execution.

As AI agents move toward practical use, the need for security and isolation keeps growing, and this tool is drawing attention as a simple yet practical approach.

---