Three things caught my attention. Apple posted a job listing that tells you more about its AI strategy than any press release. Anthropic opened Claude computer use to workplaces. And Tencent published the full Hy4 weights.
1. Apple Is Building a Mac AI Agent While Adding Friction for Third Parties
Bloomberg reported that Apple recently posted a job listing for a senior engineering manager. The listing is unusually specific about what the team is building.
The job description, as reported by TechCrunch, reads: “Imagine an AI agent that can use a computer the way a capable person does. It browses the web, works across apps, runs commands, writes and debugs code, and calls services, and it chooses the right tool for each step.” The agent is supposed to move “fluidly between tools” and “carry out multi-step tasks across apps, browsers, command lines, code, and APIs.” The listing ends with a phrase Apple customers will recognize: the agent does “all this privately and with the polish people expect from Apple.”
Apple did not respond to TechCrunch’s request for comment.
The same Bloomberg report notes Apple is tightening access for third-party agents on Mac. macOS 27.3 will add warnings that discourage users from granting full access to third-party agents. In October, Apple had already notified developers it planned to introduce additional controls around this. That is a meaningful pattern: Apple is building its own agent capability while adding friction for Anthropic, OpenAI, and others doing the same thing on its platform.
In January, Apple and Google announced that the next generation of Apple Foundation Models will be based on Google’s Gemini models, powering features including a more personalized Siri. We covered that deal in Apple Picks Google Gemini for Siri 2.0. An in-house computer-use agent is a different category: Siri handles assistant queries, while a computer-use agent operates your machine autonomously.
Why it matters. The job posting itself is evidence, not a confirmed product. Apple has not announced anything, Bloomberg says progress on the in-house agent “remains in the early stages,” and job listings sometimes reflect experiments that go nowhere. But the direction is clear. The restrictions on third-party agents in macOS 27.3, combined with Apple building in the same space, look like a platform strategy to me: build the agent yourself, then make it harder for rivals to get the same access.
Find out more: Apple may be developing its own AI agent that runs your Mac (TechCrunch)
2. Anthropic Opens Claude Computer Use to Team and Enterprise Plans
Anthropic says Claude computer use is now in beta on Team and Enterprise plans, off by default. Organizations get access on a rolling basis. Claude can see a user’s screen and work inside desktop apps, clicking and typing on their behalf. It runs on Claude Desktop for macOS 14 or later and Windows. It can also run from Claude on the web or Claude Mobile, if the organization’s Computer access setting is on and a computer has Claude Desktop open.
Owners and Primary Owners turn it on from Organization settings under Capabilities. HIPAA-ready configurations, organizations using Zero Data Retention for Claude Code, those with customer-managed encryption keys, and accounts with FERPA or K-12 settings cannot turn it on.
The support article, updated today, includes security caveats that IT teams should read before enabling this. Claude has no sandbox between its actions and a user’s applications. Content on the screen, including web pages and email, could cause Claude to take unintended actions. Claude is also not restricted to the folders a user has connected. Through approved apps, it can open, change, or delete files outside those folders. Claude takes screenshots to navigate, and those screenshots are saved with the chat. The support article notes they can include personal or confidential information.
Anthropic expanded Claude’s workplace tools on November 10, launching Claude Sheets and Whiteboards in beta and making Claude Data generally available. Computer use is a larger step because it goes beyond Claude’s own interface and into everything running on the desktop.
Why it matters. The no-sandbox note is the one that matters most. An enterprise deployment where Claude can read email, open files outside approved directories, and act on anything visible on screen is a wide attack surface. I would want to see a specific list of approved apps reviewed carefully before enabling this, and a clear policy for what happens when Claude encounters something sensitive on screen. The gradual rollout gives IT teams time to plan.
Find out more: Set up computer use for Team and Enterprise plans (Anthropic)
3. Tencent Publishes Full Hy4 Weights: 770B Parameters, Multimodal, New License

Image: Tencent
Tencent has open-sourced the full Hy4 weights on Hugging Face. The model is a mixture-of-experts architecture with 770 billion parameters in the language backbone, activating 49 billion per token. The context window is 1,048,576 tokens. It is multimodal: it handles text, image, and video through a 1.23-billion-parameter vision encoder.
The architecture detail: 78 layers total, with one standard dense feedforward layer at the front and 77 layers using mixture-of-experts feedforward networks. Each MoE layer has 256 routed experts and one shared expert. Every token activates the top 8 routed experts plus the shared one. Hy4 also includes DFly2, a 3.3-billion-parameter speculative decoder that reuses the main model’s KV cache to support long-context speculative decoding at what Tencent calls “a low, constant cost.” The model card includes vLLM and SGLang deployment instructions.
The license is tencent-hy-license. That is different from Hy4-preview, which shipped under Apache 2.0. The commercial terms matter for anyone building on this. Anyone evaluating the model for commercial use should read the full terms before going further.
Tencent’s model card lists capabilities including “Vision and computer use” and “Frontend and game development” (Source). Hy4-preview is listed on Hugging Face as a text generation model; Tencent says Hy4 adds vision capabilities.
For comparison, Mistral released a 1-trillion-parameter multimodal model with open weights on October 29.
Why it matters. A 770B multimodal open-weight model from a major Chinese technology company is a significant release. The practical question is what the tencent-hy-license allows. If the terms are more restrictive than Apache 2.0, researchers and startups building on open weights will need to evaluate carefully. The model itself is technically impressive. The license is what determines whether the open-source community can actually do what they want with it.
Find out more: Tencent Hy4 on HuggingFace

