AI News for March 19, 2026 — 7min.ai

[NEWS] Xiaomi releases MiMo-V2-Pro, a 1T-parameter model rivaling GPT-5.2 at a fraction of the cost: Xiaomi launched MiMo-V2-Pro, a 1-trillion parameter foundation model with benchmarks approaching GPT-5.2 and Opus 4.6 at roughly one-sixth the cost via API.
[NEWS] MiniMax releases M2.7, a 'self-evolving' proprietary model that automates its own RL training: Chinese AI startup MiniMax released M2.7, a proprietary reasoning model designed for AI agents and third-party tools like Claude Code and OpenClaw.
[NEWS] Apple blocks vibe-coding apps Replit and Vibecode from App Store updates: Apple is preventing vibe-coding apps including Replit ($9B valuation) and Vibecode from publishing App Store updates, citing guideline 2.5.2 that prohibits apps from executing dynamically generated code.
[RESEARCH] Anthropic publishes largest-ever qualitative AI study: 81,000 Claude users across 159 countries: Anthropic released findings from 80,508 Claude users interviewed across 159 countries in 70 languages using its AI interviewer tool — believed to be the largest multilingual qualitative study ever conducted.
[TOOLS] Perplexity launches Comet AI browser for iPhone with built-in search assistant: Perplexity released its Comet AI browser for iPhone, expanding the desktop app to iOS with built-in AI search, voice mode, and a web-based task assistant that can summarize emails, compare prices, and complete actions across sites.
[TOOLS] Google Labs turns Stitch into a full AI design platform with 'vibe design': Google Labs upgraded Stitch from a simple design tool into a full AI-powered software design platform that generates user interfaces from natural language prompts.
[TOOLS] Midjourney V8 launches with 5x faster generation and native 2K resolution: Midjourney shipped V8 for community testing, claiming roughly 5x faster image generation.
[NEWS] UK government drops plan to let AI companies use copyrighted work without permission: Technology Secretary Liz Kendall reversed the UK government's preferred option on AI copyright reform, which would have allowed tech companies to train on copyrighted work unless creators opted out.
[NEWS] Pentagon plans to let AI companies train models on classified military data: The US Department of Defense is setting up secure environments where AI companies like OpenAI and xAI can train models directly on classified data, a first.
[NEWS] OpenAI launches 'Parameter Golf': build the best LLM in 16MB for $1M in prizes: OpenAI is running an open research competition challenging developers to build the best language model under tight constraints: weights and training code combined must stay under 16MB, and training can take no longer than 10 minutes on eight H100 GPUs.

About 7min.ai · Subscribe to newsletter