One-Line Summary

EU mandates open access to parts of Google’s Android and Search while open-weight models and “physical AI” moves expand the paths teams can adopt AI.

Big Tech

Google to open Android and Search access after EU ruling

EU regulators detail changes that require Google to help rival AI assistants and search engines access its services, including opening up 11 Android features to competitors and sharing certain search optimization data with AI chatbots such as OpenAI’s, subject to anonymization and security review; users in Europe are expected to benefit starting July 2027 with the next Android iteration. ¹

Google pushes back, arguing the decisions risk weakening privacy and device safeguards for millions of Europeans, and says it has proposed alternatives that meet the Digital Markets Act’s goals without compromising security. The company warns that exposing private searches to unfamiliar providers could harm user privacy and business confidentiality. ²

For users and teams, this points to more choice in assistants and integrated actions on Android and in Search. Google is also rolling out features to link more third‑party services directly in Search’s AI Mode, showing how app‑to‑Search connections could work when multiple assistants are in the mix. ³

What to watch: the Commission’s final security criteria for accessing Android features, how anonymization is enforced for shared search data, and which rival assistants move first on Android when the July 2027 window opens. ¹

Apple Intelligence registers for use in China

China’s cyberspace regulator says Apple’s on‑device generative AI service, Apple Intelligence, is registered for use on iPhones in the country; Alibaba says its Qwen model will be integrated across iOS, iPadOS, macOS and visionOS in China, and Baidu confirms collaboration. No launch date is provided. ⁴

For companies distributing apps in China, Apple’s developer guidance highlights additional compliance fields (e.g., ICP filing numbers and other permits) that appear on App Store product pages in mainland China. Teams should confirm that required identifiers and documentation are configured correctly in App Store Connect before shipping AI‑enhanced features. ⁵⁶

Nvidia pushes ‘physical AI’ in Japan and new agent tooling

Nvidia introduces Cosmos 3 Edge, a world model for robots and vision agents, and expands its physical AI ecosystem in Japan with partners like Fujitsu, Hitachi and Kawasaki; it also spotlights healthcare collaborations through BioNeMo for drug discovery. The moves underscore Nvidia’s focus on AI that perceives and acts in real environments. ⁷

On the tooling front, Nvidia details nanousd‑labs, which uses AI agents to generate lightweight OpenUSD runtimes from the standard itself, and a blueprint showing how to integrate context‑aware video AI agents with enterprise workflows using NemoClaw—turning video analysis into structured reports and tickets. This is relevant to ops, compliance, and content teams that process large video libraries. ⁸⁹

Industry & Biz

Moonshot debuts Kimi K3 to challenge top models

VentureBeat reports that Beijing‑based Moonshot AI releases Kimi K3, a 2.8‑trillion‑parameter open‑weight model with a 1‑million‑token context window, and plans to publish full weights on July 27; the company also claims competitive results versus leading proprietary systems and offers API access now. ¹⁰

TechCrunch, citing the Financial Times, previously reported K3 would target performance parity with Anthropic’s latest models, use 2–3 trillion parameters, and arrive alongside fundraising that could value Moonshot at $31.5 billion. Treat the valuation and timing as reporting from those outlets, not company-confirmed figures. ¹¹¹²

Why it matters: open‑weight options that approach frontier performance can reduce per‑request costs, ease data‑control concerns, and give enterprises another model to fine‑tune for internal tasks—especially for teams hesitant to send sensitive prompts to closed APIs. VentureBeat notes pricing at $3 per million input tokens and $15 per million output tokens, with discounted cached input. Try the chat experience at kimi.com to gauge fit for your use cases. ¹⁰

Thinking Machines releases Inkling, a 975B-parameter open-weights multimodal model

Thinking Machines—founded by former OpenAI leaders—launches Inkling under the permissive Apache 2.0 license, with 975 billion total parameters (41 billion active), native text‑vision‑audio reasoning, a 1‑million‑token context window, and a design that emphasizes controllable “thinking effort.” Weights are available and the model can be customized on the company’s Tinker platform. ¹³

External write‑ups note strong but sub‑state‑of‑the‑art scores on some coding and multimodal benchmarks, with particular strengths in voice understanding and agentic workflows. The company itself frames Inkling as a broad, customizable base rather than a benchmark champion. ¹⁴¹³

For enterprises, the Apache 2.0 license provides legal clarity to download, modify and commercialize derivatives. The model card emphasizes pairing open weights with defense‑in‑depth (e.g., external moderation) when deploying safety‑sensitive features—especially after fine‑tuning. ¹³

What This Means for You

If you operate in the EU or serve EU Android users, expect more choice in voice assistants and tighter scrutiny on how search data is shared and anonymized. Product teams can plan scenarios where a non‑Google assistant handles bookings or local queries, while legal and security teams review access rules ahead of July 2027. ¹

Open‑weight momentum gives you new build‑vs‑buy levers. Kimi K3 aims to bring “frontier‑class” performance to open weights, while Inkling offers permissive licensing and multimodality for on‑prem or VPC deployments—useful when data governance or cost predictability matters. Pilot with a narrow workflow before scaling. ¹⁰¹³

If you have customers or distribution in China, Apple Intelligence’s registration is a signal to align your iOS roadmap with local compliance. Check your App Store Connect fields (e.g., ICP, content permits) so your AI features and metadata are ready to go when Apple’s rollout timeline becomes clear. ⁴

For operations, safety, or media teams, Nvidia’s blueprints show a practical pattern: turn video into structured reports and downstream tickets. Even without robotics, that “analyze then act” flow can help with audits, content QA, incident reviews, and training libraries. ⁹

Action Items

1. Test Moonshot Kimi K3 chat: Create an account on kimi.com and run a few of your real documents or tasks to compare reasoning quality and latency against your current model.
2. Try Inkling in Tinker Playground: Use the playground to probe multimodal prompts (image + text, or audio + text) and see how controllable thinking effort affects cost and quality.
3. Prepare China compliance in App Store Connect: Verify ICP and other required fields are correctly populated on your product pages to avoid launch delays for AI‑enhanced apps.
4. Map an EU Android assistant scenario: List one user flow (e.g., ride‑hailing or restaurant search) you’d hand to a non‑Google assistant and outline the privacy and security checks your org would need.
5. Draft a one‑day video‑to‑ticket pilot: With your IT or data team, outline a small trial that summarizes a training or audit video into a timestamped report and a Jira (or equivalent) ticket.