SpaceXAI Open-Sources Grok Build and Resets All Quotas

This edition of AI Highlights covers 22 stories across model releases, developer tools, products and industry news. Here's the rundown.

Top Story

SpaceXAI Open-Sources Grok Build, Resets Quotas and Disables Default Data Retention

SpaceXAI announced it has open-sourced Grok Build, its coding agent and terminal interface, releasing the Rust source covering the agent loop, tool implementations, terminal UI and extension system.

SpaceXAI also reset usage quotas for all users, switched Grok Build's data-retention default to off, and deleted all previously retained coding data.

SpaceXAI said early beta builds enabled data retention by default for users with non-zero retention, but stressed that users could always disable data upload from the command line and that this choice was respected.

Users can compile the source and configure their own local inference to run Grok Build.

Sources: github.com · x.ai

Model Releases

Thinking Machines Lab Releases the Multimodal Open Model Inkling

Thinking Machines Lab released Inkling, a multimodal mixture-of-experts model, with full weights available and fine-tuning supported on the Tinker platform.

The model has 975B total parameters and 41B active parameters, supports up to a 1M context window, natively handles text, image and audio, and lets you trade off performance against cost via adjustable inference compute.

Weights are available on Hugging Face in both the original form and an NVFP4 build tuned for NVIDIA Blackwell, with local deployment supported on frameworks such as SGLang and vLLM, plus APIs from several third-party providers.

The lightweight variant Inkling-Small is currently preview-only; full weights will ship after testing.

Sources: thinkingmachines.ai · huggingface.co · tinker.thinkingmachines.ai

ByteDance Seed Upgrades Its Model: 1M Context and a Cost Promotion

The ByteDance Seed team announced three major upgrades to the Doubao-Seed-Evolving model.

The model now supports a 1M ultra-long context, able to handle large code repositories and long documents.

Its long-horizon task ability scored above Doubao-Seed-2.1-pro in developer crowd-testing, while consuming fewer tokens and using more concise tool-call rounds.

An "experience reward" program currently offers a 20% discount on usage cost.

The model is available on the Volcano Agent Plan, with monthly bundles starting at a limited-time 9.9 RMB.

Sources: mp.weixin.qq.com · ark.volcengine.com

OpenAI Releases GPT-Red, an Internal Automated Red-Teaming Model

OpenAI released GPT-Red, an internal automated red-teaming model used to find prompt-injection vulnerabilities in its models.

Trained via self-play reinforcement learning, GPT-Red can break almost all internal and production models including GPT-5.5, and has been used to adversarially train GPT-5.6 Sol.

According to official figures, compared with the best production model from four months earlier, GPT-5.6 Sol reduced its failure rate on the hardest direct prompt-injection benchmark by 6x.

To prevent leaking malicious capability, GPT-Red is isolated from deployed models and remains internal-only for now.

Sources: openai.com · x.com

Qwen Releases the Real-Time Voice Model Qwen-Audio-3.0-Realtime

Qwen officially released Qwen-Audio-3.0-Realtime, a real-time voice conversation model, now available on Alibaba Cloud Bailian.

It comes in a stronger-reasoning Plus version and a faster Flash version, delivering millisecond-level responses without sacrificing reasoning depth, and supports full-duplex interaction — listening while speaking and interruptible at any time.

Unlike traditional voice models that rely on explicit instructions, it can call external tools such as MCP and APIs on its own without explicit prompts, and the results are automatically woven into subsequent conversation memory.

It can also dynamically adjust tone and emotion to context, and the company says it scored high or achieved SOTA on several benchmarks.

Sources: mp.weixin.qq.com · bailian.console.aliyun.com

Xiaomi Open-Sources the Embodied Generation Model Xiaomi-Robotics-U0

Xiaomi released and open-sourced Xiaomi-Robotics-U0, a 38B-parameter autoregressive world foundation model.

The model uses a unified multimodal autoregressive framework covering four core tasks: embodied scene generation, embodied transfer, embodied video generation and general vision-language generation.

Official data shows it ranked first among 126 models on the WorldArena benchmark, with real-robot OOD task completion improving by over 26% on average.

Code and part of the model weights are open now; video-generation weights are coming soon.

Sources: robotics.xiaomi.com · huggingface.co

Dev Tools & Ecosystem

Claude Code Artifacts Now Support Calling MCP Connectors

Claude Code artifacts can now call MCP connectors.

This lets users build dashboards and apps that fetch information and take actions on demand for each viewer.

The feature is currently available only on the Pro, Max, Team and Enterprise subscription plans.

It is not supported for publicly shared artifacts.

Sources: x.com

Cursor Doubles Cursor-Model Usage Across All Plans

Cursor announced it has doubled the included Cursor-model usage across all subscription plans.

The change applies to every plan, giving users more access to Grok 4.5 and Composer 2.5.

The update is effective immediately.

Sources: x.com

OpenAI Launches a GPT-5.6 Share Campaign: Post for $100 in ChatGPT Credits

OpenAI launched a community share campaign for the GPT-5.6 Sol model.

Users simply post publicly on X about why they like GPT-5.6, how it compares to other models, or something they built with it.

Submit the post link and your ChatGPT account email to take part.

The reward is $100 in ChatGPT credits, roughly 2,500 credits.

Participants must have an active ChatGPT Go, Plus or Pro subscription.

Sources: switch-to-codex.openai.chatgpt.site · x.com

Products & Applications

Google Expands Gemini Spark to More Regions for AI Ultra Subscribers, With Four Updates

Google announced that Gemini Spark is now rolling out to Google AI Ultra subscribers in more countries and languages.

The personal AI agent gains four updates starting today, including the ability to open and edit Google Docs directly, read comments in Google Sheets and Slides, a speed boost of over 50%, and support for parallel processing across multiple sources.

For now these features are limited to Google AI Ultra subscribers on the web or in the app.

Sources: support.google.com

OpenAI Raises the ChatGPT Custom-Instructions Character Limit to 5,000

OpenAI has raised the character limit for ChatGPT custom instructions, and the change is now live.

Plus, Pro, Enterprise, Business and Education users can now save custom instructions of up to 5,000 characters in ChatGPT.

The previous limit for those users was 1,500 characters.

The extra room lets users tailor ChatGPT's response style and behavior in more detail.

Sources: help.openai.com

Alipay and OPPO Reach a Cross-Device Agent Interconnect Partnership

Alipay and OPPO reached a cross-device agent interconnect partnership, linking Alipay's agent "Abao" with OPPO's phone assistant "Xiaobu."

The feature is now live on OPPO phones. OPPO users can invoke, via voice through Xiaobu, nearly 200 life services from Abao's first integration batch — without opening the Alipay app.

Abao automatically matches and pre-handles preliminary steps in the background, with some services running silently. Key payment and authorization steps must be confirmed by the user manually.

The feature requires upgrading ColorOS, Xiaobu Assistant, Alipay and the intent-execution framework to specified versions or above.

Sources: mp.weixin.qq.com

JD's AI Agent Integrates With Tencent Yuanbao's Mini-Program Ecosystem for In-Chat Ordering

JD and Tencent jointly announced that JD's AI Agent has integrated with Tencent Yuanbao's mini-program ecosystem.

When a user queries product info in natural language inside the Yuanbao app, Yuanbao displays results as product cards. Tapping a card opens the JD shopping mini-program to complete ordering, after-sales and the full-chain service — all without leaving the conversation or switching apps.

JD's AI Agent is the first e-commerce vertical partner to integrate with Tencent Yuanbao, providing it with all-category product information.

Sources: mp.weixin.qq.com

OpenAI and Work Louder Launch the Codex Micro Keyboard

The Codex Micro keyboard, co-designed by OpenAI Supply Co. and Work Louder, is now open for online order.

It is priced at $230 and expected to ship on July 24. Thomas Ricouard said this is a limited-edition run.

The keyboard features RGB lighting that shows agent status in real time, plus a knob for adjusting reasoning level, a joystick for triggering common workflows, and dedicated command keys.

It is compatible with Mac and Windows and supports Bluetooth and USB-C.

Sources: openai.com · x.com

Tech & Insights

Arena.ai Introduces a New Ranking Combining Human Preference and Factuality

Arena.ai officially introduced a new model ranking in its text and search arenas that combines human preference with factual accuracy.

It is currently offered as a non-default toggle in the leaderboard UI, with a default factuality weight of 25%. To power the ranking, the Arena team labeled over 2 million claims generated by LLMs in real conversations — over 1.3 million from the text arena and over 700,000 from the search arena.

In the text arena, key changes after enabling the factuality weight include: Claude Fable 5 slipping slightly to 2nd; GPT-5.5 jumping 13 places to 7th, the biggest riser; and Muse Spark falling from 7th to 20th, the biggest drop.

Sources: arena.ai

Industry News

China's CAC Publishes 7 On-Device Mobile AI Filings; Apple Intelligence–Alibaba Qwen Partnership Confirmed

Chinese authorities released filing information for 7 on-device generative AI services for mobile.

Among them, Apple's Apple Intelligence completed its filing on 2026-07-08, applicable to iPhones.

Alibaba confirmed that Alibaba Qwen will be integrated into Apple Intelligence as an AI capability.

Chinese users will be able to use Qwen's text and image understanding and content generation directly on Apple devices without switching apps.

Sources: cac.gov.cn · mp.weixin.qq.com

CNKI Announces Removal of Papers That List AI as an Author

CNKI issued a statement that it has removed papers listing artificial intelligence as an author, making clear it does not accept AI tools being named as paper authors.

CNKI noted that AI lacks legal personhood and cannot bear academic responsibility, so crediting it as an author violates the law and publishing ethics.

Authors who use AI tools must clearly state so in the methods or acknowledgments section.

Papers that violate this principle may be denied publication and distribution by CNKI.

Sources: piccache.cnki.net

Anaconda Acquires the Open-Source AI Coding Agent Kilo Code

Anaconda acquired Kilo Code, an open-source AI coding agent, integrating it into its platform to cover the full AI-native development lifecycle.

Official data shows Kilo Code has over 3 million developers, dispatches nearly 10 trillion tokens per month, and its model gateway supports over 500 models.

After the acquisition, Kilo Code remains available to individual builders, teams and organizations, usable in tools such as VS Code, the CLI and JetBrains.

Sources: anaconda.com

ModelBest Closes a New Funding Round; Valuation Said to Exceed 20 Billion RMB

ModelBest announced it has closed a new funding round, with cumulative financing exceeding 5 billion RMB in the first half of 2026.

The company says its valuation now exceeds 20 billion RMB, making it the highest publicly valued unicorn in on-device intelligence.

Investors in this round include national funds, state-owned enterprises, automakers and well-known financial investors.

Its on-device large models have reached mass production in multiple vehicle models from Changan, SAIC and Geely.

Sources: zhidx.com

Outlook & Rumors

Kimi Appears to Tease K3 With a Video Full of the Number 3

Kimi, from Moonshot AI, released a teaser video on its official social accounts in which the number "3" appears repeatedly.

The "3" recurs in many forms throughout the video, widely read as hinting that the next-generation LLM Kimi K3 is coming soon.

The company has not disclosed K3's launch timing, model specs or capabilities; the teaser alone has sparked interest and speculation.

Sources: x.com

Elon Musk Says 𝕏 Will Open-Source Its Entire Codebase After a Security Review

Elon Musk said that after completing a security-vulnerability review, 𝕏 will open-source its entire codebase, with no exceptions.

To ensure transparency, 𝕏 will also invite third-party reviewers to inspect the running system and confirm the open-sourced code matches what is actually running.

The plan awaits completion of the security review before it can proceed.

Sources: x.com

Nubia Says the World's First AI-Agent Phone Will Debut at WAIC 2026

Nubia announced that the world's first AI-agent phone will make its official debut at WAIC 2026.

The product emphasizes a bolder design statement and says it will open a new chapter for AI-agent phones.

The on-site showcase is confirmed for July 17–20.

Sources: mp.weixin.qq.com