# AI News — July 17, 2026: Kimi K3's 2.8T Weights Drop, EU Cracks Open Google's Android Moat > Source: > Published: 2026-07-17 06:00:09+00:00 Good morning. It’s a big day for open weights out of China — Moonshot’s Kimi K3 is here at 2.8 trillion parameters, the largest open-weight model ever released, and early testers say it’s playing in the same league as Opus 4.8 and GPT-5.6 Sol. Elsewhere, the EU is forcing Google to pry open Android and Search for AI rivals, Apple Intelligence gets its long-awaited China green light, and a self-reported near-perfect ARC-AGI-3 score is drawing very justified skepticism. **Kimi K3 lands.** Moonshot has released [Kimi K3](https://www.kimi.com/blog/kimi-k3), a 2.8T-parameter open model with native vision and a 1M-token context window, featuring new tricks like Kimi Delta Attention and Attention Residuals that Moonshot says give it ~2.5× better scaling efficiency than K2. Weights are promised by July 27, 2026. The [Financial Times, via TechCrunch](https://techcrunch.com/2026/07/16/moonshots-upcoming-kimi-3-is-expected-to-close-the-gap-with-anthropics-opus-4-8/), reports Moonshot is raising at a $31.5B valuation, up from $20B two months ago. As a proof of concept, Moonshot had K3 autonomously design a working chip on a 45nm process in a 48-hour run using open-source EDA tools. The [HN thread](https://news.ycombinator.com/item?id=48935342) is broadly positive on quality — early Artificial Analysis numbers put it near Sol Max on cost-per-task ($0.94 vs $1.04) and above Opus 4.8 on most benchmarks — but the $3/$15 per million token pricing raised eyebrows for an open Chinese model. One commenter framed the broader Chinese lab pattern as classic commoditize-your-complement: drive intelligence toward zero cost, sell the infrastructure underneath. It’s a reasonable read. **EU tells Google to open Android and Search.** [The Verge reports](https://www.theverge.com/policy/966438/eu-google-android-ai-interoperability-search-data-dma) that under the Digital Markets Act, Google must give rival search engines and AI assistants comparable access to Android system features (deadline: July 2027) and to Google Search data (January 2027). Non-compliance risks fines of up to 10% of global revenue. The practical implication is that Gemini’s default-assistant advantage on Android and Google Search’s data moat both take structural hits in Europe. **Apple Intelligence cleared for China.** [Apple got CAC approval](https://techcrunch.com/2026/07/16/apple-intelligence-approved-for-launch-in-china-with-alibabas-qwen-ai/) to launch Apple Intelligence in China, with Alibaba’s Qwen as the primary model partner and Baidu also confirmed. DeepSeek and ByteDance integrations are reportedly being explored. Greater China did $20.5B in Q2 revenue for Apple, up 28% year-over-year, so getting this over the line matters more than the tepid AI branding suggests. **Schema Harness claims ~99% on ARC-AGI-3, sort of.** A group calling itself [Schema Harness](https://schema-harness.github.io/) says it hit ~99% RHAE on the ARC-AGI-3 Public set by wrapping Opus 4.8 and Sol in a “physicist-style” reasoning scaffold that turns raw game observations into tracked objects and mechanisms. Two catches, both flagged loudly on [HN](https://news.ycombinator.com/item?id=48935905): games scoring below 80 are rerun with a second model and the higher score kept (a pass@n in a trench coat), and the numbers are on the public set with no held-out submission and no open-sourced harness. Chollet’s original warning about hard-coded priors sneaking into ARC benchmarks feels relevant here. **NVIDIA’s Nemotron 3 Embed tops RTEB.** NVIDIA [released Nemotron 3 Embed](https://huggingface.co/blog/nvidia/nemotron-3-embed-wins-rteb), an open-weight embedding family aimed at enterprise RAG, with the 8B flagship taking #1 on the RTEB multilingual leaderboard as of July 15. There are 1B variants for production and an NVFP4 quantized option tuned for Blackwell. 32k context, multilingual, code retrieval included. **LM Studio ships Bionic.** [Bionic](https://lmstudio.ai/blog/introducing-lm-studio-bionic) is LM Studio’s new agent for coding and document work with open models, running locally or through a new zero-retention cloud service, with offline voice transcription via Mistral’s Voxtral. Early users on [HN](https://news.ycombinator.com/item?id=48939662) say the Codex-style UI is pleasant and it works well with GLM 5.2 and Kimi K2.7. The friction point: both LM Studio and Bionic are closed-source despite the open-model positioning, and the cloud pivot has some users nervous about where this is headed. **NotebookLM becomes Gemini Notebook.** Google [renamed NotebookLM to Gemini Notebook](https://blog.google/innovation-and-ai/products/gemini-notebook/notebooklm-gemini-notebook/), tightening integration with the Gemini app and Search, and adding a secure cloud sandbox for native code execution — available now for AI Ultra, coming to Pro. The product is at 30 million users and 600,000 organizations. Reaction on [HN](https://news.ycombinator.com/item?id=48936451) was mostly the expected Google-graveyard cynicism, with one commenter noting the other labs are shipping models while Google renames things. **Detecting LLM text with old-fashioned ML.** A developer [built a classical ML detector](https://blog.lyc8503.net/en/post/llm-classifier/) that reaches ~85% single-sentence accuracy by leaning on the statistical fingerprints current LLMs leave behind, similar in spirit to how commercial AI checkers likely operate. The [HN discussion](https://news.ycombinator.com/item?id=48936880) is skeptical about longevity — text isn’t information-dense enough to carry robust provenance signals — though one commenter mentions a small encoder-only transformer hitting 99.81 AUROC on RAID-bench, suggesting the ceiling is much higher than 85%. **Ring-Zero scales verifiable-reward RL to 1T parameters.** Ant Group researchers [describe Ring-Zero](https://arxiv.org/abs/2607.12395), a “zero RL” training run (verifiable rewards, no human labels) at trillion-parameter scale, reporting emergent self-verification, parallel reasoning, and structured formatting along with stabilization tricks like clipped importance sampling. [HN commenters](https://news.ycombinator.com/item?id=48940603) were less impressed with the LLM-as-judge evaluation loop, calling it incestuous, and one noted the ongoing absurdity of burning gigawatts to inch toward what a 20-watt brain does for free. That’s the morning. Kimi K3 will dominate benchmark chatter through the weekend — the real test comes when the weights actually drop on the 27th and people can prod them without paying $15 per million output tokens.