AI News: June 14, 2026

1. Moonshot AI Releases Open-Weight Kimi K2.7 Code at a Fraction of Rivals’ Pricing

Moonshot AI. The Chinese lab released Kimi K2.7 Code, an open-weight mixture-of-experts model with 1 trillion total parameters, 32 billion active per token, and a 256,000-token context window, available on Hugging Face under a modified MIT license. The model trails Western frontier models on standard coding benchmarks but beats Claude Opus 4.8 on the agent-focused MCPMark Verified test, while pricing at $0.95 per million input tokens and $4.00 per million output tokens, up to 12 times cheaper than Claude Fable 5 on output. Source

2. Tsinghua Researchers Release “Count Anything” Object-Counting Model

Tsinghua University. Researchers released Count Anything, a model that counts and labels objects across image types from satellite imagery to medical scans using text prompts, combining bounding boxes for large objects with point placement for small dense targets. Built on Meta’s SAM3 with adapter components and trained on a custom 220,000-image dataset called CLOC, it miscounts by roughly nine objects per category on average, less than half the error of the best alternatives, with code available on GitHub. Source

3. KPMG Pulls Agentic AI Report After Fabricated Client Claims Surface

KPMG. The consulting firm withdrew its report “Redefining excellence in the age of agentic AI” after organizations including UBS, the UK’s National Health Service, Swiss Federal Railways, and Transport for London said statements about their AI usage were untrue or misleading. Research group GPTZero identified the inaccuracies as apparent AI hallucinations, in an episode where a report about AI usage was itself undermined by unverified AI-generated content. Source

4. OpenAI Served Subpoena in State Attorneys General Investigation

State attorneys general. A coalition led by New York’s attorney general served OpenAI with a subpoena seeking documents on its advertising practices, user engagement and retention, model behavior, handling of consumer and health data, and treatment of minors and seniors. OpenAI said it takes the concerns seriously and pointed to protective features for minors such as age prediction and parental tools, though the specific states involved were not disclosed. Source

5. Meta Reportedly Unwinds $2B Manus Acquisition After Beijing Divestiture Order

Meta. The company has reportedly begun operationally separating from Manus, the Chinese-founded AI startup it acquired for $2 billion in December 2025, after Beijing issued a divestiture order on national security grounds citing potential technology export-control and foreign-investment violations. Meta is cutting Manus off from internal systems while the startup’s co-founders explore raising $1 billion to reclaim the company, potentially via a Chinese joint venture aimed at a Hong Kong listing. Source

6. India Debates Sovereign AI Push After US Forces Anthropic Model Shutdown

India tech sector. The US government directive forcing Anthropic to suspend its newest models prompted Indian technology and policy leaders to call for accelerated domestic AI capability, with figures arguing the episode underscores the geopolitical risk of depending on American frontier models. Proposals included a 500 billion rupee ($5 billion) annual AI fund and a 2 trillion rupee credit-guarantee program for infrastructure, alongside calls to embrace smaller Indian and Chinese open-source models. Source

7. SkillOpt Research Trains Markdown “Skills” to Boost LLM Agents by 23 Points

Microsoft and university researchers. A new arXiv paper introduces SkillOpt, a method that optimizes plain-text instruction documents for AI agents as if they were model weights, using a separate optimizer model to edit a Markdown file while the base model stays frozen. The approach delivered roughly 23-point improvements on GPT-5.5 across six benchmarks, outperforming handwritten skills and prior methods like Trace2Skill and TextGrad while keeping skills compact at 300 to 2,000 tokens and transferable across models. Source